File size: 89,867 Bytes
57e71f8 7211e63 57e71f8 7211e63 5719ec3 57e71f8 7211e63 5719ec3 57e71f8 a144947 57e71f8 7211e63 5719ec3 7211e63 5719ec3 57e71f8 7211e63 5719ec3 7211e63 57e71f8 a0bdb90 7211e63 583fea5 57e71f8 5719ec3 57e71f8 a0bdb90 7211e63 57e71f8 a144947 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 f7f5220 57e71f8 5719ec3 57e71f8 a144947 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 7211e63 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 533611a 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 57e71f8 5719ec3 14a2669 5719ec3 57e71f8 a0bdb90 7211e63 57e71f8 7211e63 57e71f8 a0bdb90 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 a0bdb90 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 a0bdb90 7211e63 57e71f8 a0bdb90 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 7211e63 57e71f8 5719ec3 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 a0bdb90 57e71f8 7211e63 57e71f8 7211e63 57e71f8 5719ec3 57e71f8 7211e63 57e71f8 7211e63 57e71f8 5719ec3 57e71f8 7211e63 57e71f8 5719ec3 7211e63 5719ec3 7211e63 5719ec3 7211e63 14a2669 7211e63 14a2669 7211e63 5719ec3 14a2669 a144947 14a2669 a144947 14a2669 a144947 14a2669 a144947 14a2669 a144947 14a2669 a144947 14a2669 7211e63 57e71f8 7211e63 5719ec3 57e71f8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
CyberSOCEnv β Enterprise Cybersecurity Operations Center Environment.
Implements the OpenEnv Environment interface for a deterministic SOC
incident response simulation on a 500-node enterprise network.
The agent receives SIEM/EDR alerts, queries hosts, runs forensics,
isolates segments, blocks IOCs, kills processes, and submits a
containment plan β all while minimizing business downtime.
"""
from __future__ import annotations
import copy
import random
import uuid
from typing import Any, Callable, Dict, List, Optional
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
try:
from ..models import (
SOCObservation,
SOCActionWrapper,
SOCState,
Alert,
NetworkTopology,
ForensicsResult,
TimelineEntry,
QueryHost,
IsolateSegment,
BlockIOC,
RunForensics,
KillProcess,
SubmitContainmentPlan,
CorrelateAlerts,
EnrichIOC,
ScanHostVulnerabilities,
TerminatePID,
CreateFirewallRule,
QuarantineFile,
RedActionWrapper,
LateralPivot,
DeployPayload,
EvadeDetection,
PassTurn,
RED_ACTION_TYPES,
)
except ImportError:
from models import (
SOCObservation,
SOCActionWrapper,
SOCState,
Alert,
NetworkTopology,
ForensicsResult,
TimelineEntry,
QueryHost,
IsolateSegment,
BlockIOC,
RunForensics,
KillProcess,
SubmitContainmentPlan,
CorrelateAlerts,
EnrichIOC,
ScanHostVulnerabilities,
TerminatePID,
CreateFirewallRule,
QuarantineFile,
RedActionWrapper,
LateralPivot,
DeployPayload,
EvadeDetection,
PassTurn,
RED_ACTION_TYPES,
)
from .tasks import get_task, build_network
from .graders import grade_episode
from .threat_graph import (
ThreatGraph,
HostNode,
ProcessNode,
IOCNode,
VulnerabilityNode,
AlertNode,
Edge,
)
class ActionMiddleware:
"""Pre-flight validation for SOC actions.
Detects phase violations (action out of order) and graph-ungrounded actions
(action references an entity not yet discovered in the ThreatGraph).
Returns None if the action is valid, or an error dict otherwise.
"""
def validate(
self,
current_phase: str,
action_type: str,
args: Dict[str, Any],
graph,
) -> Optional[Dict[str, str]]:
# Phase violation: plan submission before any investigation
if action_type == "submit_containment_plan" and current_phase == "triage":
return {
"error_type": "PHASE_VIOLATION",
"message": "submit_containment_plan requires investigation phase first",
}
# Graph-groundedness: IOC must be discovered before enrichment
if action_type == "enrich_ioc":
ioc_val = args.get("ioc_value", "")
if ioc_val and graph is not None and ioc_val not in graph.iocs:
return {
"error_type": "GRAPH_FAILURE",
"message": f"IOC '{ioc_val}' not in threat graph; receive an alert or run forensics first",
}
# Graph-groundedness: host must be known before vulnerability scan
if action_type == "scan_host_vulnerabilities":
hostname = args.get("hostname", "")
if hostname and graph is not None and hostname not in graph.hosts:
return {
"error_type": "GRAPH_FAILURE",
"message": f"Host '{hostname}' not in threat graph; run query_host first",
}
# Emergency isolation gate: allow early isolate_segment only when a critical
# alert proves an active threat on the targeted subnet/host; otherwise penalise
# the panic as UNJUSTIFIED_EMERGENCY.
if action_type == "isolate_segment" and current_phase == "triage":
subnet = args.get("subnet", "")
target_host = args.get("target_host", "")
has_critical = False
if graph is not None:
for alert in graph.alerts.values():
if alert.severity != "critical":
continue
src = alert.source_host
if target_host and src == target_host:
has_critical = True
break
if subnet and src in graph.hosts:
host_node = graph.hosts.get(src)
if host_node and getattr(host_node, "subnet", "") == subnet:
has_critical = True
break
if not has_critical:
return {
"error_type": "UNJUSTIFIED_EMERGENCY",
"message": (
"isolate_segment during triage requires a critical-severity alert "
"on the targeted subnet/host to justify emergency response"
),
}
return None
class CyberSOCEnvironment(Environment):
"""
Deterministic SOC incident response environment.
Simulates a 500-node enterprise network under attack. The agent must
investigate alerts, contain threats, and submit a containment plan
while minimizing business downtime.
Supports concurrent WebSocket sessions (each gets own instance).
Example:
>>> env = CyberSOCEnvironment()
>>> obs = env.reset(task_id="easy")
>>> print(len(obs.alert_queue)) # Initial alerts
>>> obs = env.step(SOCActionWrapper(type="query_host", hostname="WS-042"))
"""
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(
self,
adaptive: bool = False,
neural_red_policy: Optional[Any] = None,
red_team_logger: Optional[Callable[[Dict[str, Any]], None]] = None,
fsp_mode: bool = False,
):
"""Initialize the environment (actual state set in reset).
Args:
adaptive: Legacy adaptive-adversary flag (kept for backward compat).
neural_red_policy: Optional callable for neural Red policy (legacy hook).
red_team_logger: Optional callback for recording Red decisions.
fsp_mode: When True, step() uses strict alternating turns and
step_count only increments after BOTH Blue and Red have acted.
When False (default), step(SOCActionWrapper) behaves exactly as
before β Red's PassTurn is applied automatically so existing code
and tests remain unaffected.
"""
super().__init__()
self._adaptive = adaptive
self._neural_red_policy = neural_red_policy
self._red_team_logger = red_team_logger
self._fsp_mode = fsp_mode
self._red_team_decisions: List[Dict[str, Any]] = []
self._live_requirements: Dict[str, Any] = {}
self._threat_graph = None # will be initialized on reset()
self._state = SOCState(episode_id=str(uuid4()), step_count=0)
self._network: Dict[str, List[Dict[str, Any]]] = {}
self._task_def: Dict[str, Any] = {}
self._alert_queue: List[Dict[str, Any]] = []
self._host_index: Dict[str, Dict[str, Any]] = {} # hostname -> host dict
self._plan_entries: List[Dict[str, Any]] = []
self._last_forensics: Optional[ForensicsResult] = None
self._middleware = ActionMiddleware()
self._rng = random.Random(0) # overwritten in reset()
self._pending_followup: Dict[str, bool] = {} # hostname -> responded_to
self._disruption_cost: float = 0.0 # accumulates per clean host/subnet isolated
self._discovered_iocs: set = set() # IOCs revealed via run_forensics or enrich_ioc
self._quarantined_files: set[tuple[str, str]] = set()
self._step_reward_total: float = 0.0
def _reset_rubric(self):
"""Initialize live containment requirements for dynamic grading in adaptive mode."""
import copy
self._live_requirements = copy.deepcopy(
self._task_def.get("containment_requirements", {})
)
# ===========================================================================
# reset()
# ===========================================================================
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
**kwargs: Any,
) -> SOCObservation:
"""Reset the environment for a specific task.
Args:
seed: Ignored (environment is fully deterministic).
episode_id: Optional custom episode ID.
**kwargs: Must include task_id ('easy', 'medium', or 'hard').
Returns:
Initial SOCObservation with alert queue and network state.
"""
task_id = kwargs.get("task_id", "easy")
self._rng = random.Random(hash(task_id))
self._task_def = get_task(task_id)
self._recent_actions = [] # reset stall detector
# Build deterministic network (cached per task for GRPO throughput)
if not hasattr(CyberSOCEnvironment, "_network_cache"):
CyberSOCEnvironment._network_cache = {}
cache_key = task_id
if cache_key in CyberSOCEnvironment._network_cache:
self._network = copy.deepcopy(CyberSOCEnvironment._network_cache[cache_key])
else:
self._network = build_network()
CyberSOCEnvironment._network_cache[cache_key] = copy.deepcopy(self._network)
# Build hostname index for O(1) lookups
self._host_index = {}
for subnet_name, hosts in self._network.items():
for host in hosts:
self._host_index[host["hostname"]] = host
# Inject attack chain: mark compromised hosts, add malicious processes
for threat in self._task_def["attack_chain"]:
for hostname in threat["compromised_hosts"]:
if hostname in self._host_index:
host = self._host_index[hostname]
host["status"] = "compromised"
for proc in threat["malicious_processes"]:
if proc not in host["running_processes"]:
host["running_processes"].append(proc)
# Initialize alert queue (deep copy so mutations don't affect task def)
self._alert_queue = copy.deepcopy(self._task_def["initial_alerts"])
# Reset state
eid = episode_id or str(uuid4())
self._state = SOCState(
episode_id=eid,
step_count=0,
task_id=task_id,
max_steps=self._task_def["max_steps"],
total_reward=0.0,
business_impact=self._task_def["initial_business_impact"],
contained_threats=[],
active_threats=[t["threat_id"] for t in self._task_def["attack_chain"]],
blocked_iocs=[],
isolated_subnets=[],
forensics_run=[],
killed_processes=[],
queried_hosts=[],
timeline=[],
is_done=False,
submitted_plan=False,
active_turn="blue",
)
self._plan_entries = []
self._last_forensics = None
self._reset_rubric()
self._fired_step_rewards: set = set()
self._step_reward_total: float = 0.0
self._pending_followup: Dict[str, bool] = {}
self._disruption_cost = 0.0
self._discovered_iocs: set = set()
self._quarantined_files: set[tuple[str, str]] = set()
self._red_team_decisions = []
# Initialize threat graph from task definition
self._threat_graph = ThreatGraph()
self._populate_threat_graph()
# Inject external threat-intel feed IOCs so Blue can immediately enrich/block them
# without hitting GRAPH_FAILURE (simulates acting on CISA or partner feed data).
for ioc_entry in self._task_def.get("external_intel_feed", []) or []:
if isinstance(ioc_entry, str):
ioc_value = ioc_entry
parts = ioc_entry.split(".")
if len(parts) == 4 and all(p.isdigit() for p in parts):
ioc_type = "ip"
elif len(ioc_entry) >= 32 and "." not in ioc_entry:
ioc_type = "hash"
else:
ioc_type = "domain"
elif isinstance(ioc_entry, dict):
ioc_value = ioc_entry.get("value", "")
ioc_type = ioc_entry.get("type", "ip")
else:
continue
if not ioc_value:
continue
if ioc_value not in self._threat_graph.iocs:
self._threat_graph.add_ioc(
IOCNode(ioc_value=ioc_value, ioc_type=ioc_type, confidence=0.70)
)
self._discovered_iocs.add(ioc_value)
self._last_obs_extras: Dict[str, Any] = {}
return self._build_observation(reward=0.0, done=False)
def _populate_threat_graph(self) -> None:
"""Seed the threat graph with hosts, processes, IOCs, and alerts from task_def."""
graph = self._threat_graph
# Hosts: include compromised hosts from attack chain + every host they live on
compromised_set: set[str] = set()
for threat in self._task_def.get("attack_chain", []):
for hn in threat.get("compromised_hosts", []):
compromised_set.add(hn)
for hostname in compromised_set:
host_dict = self._host_index.get(hostname)
if host_dict is None:
continue
graph.add_host(HostNode(
hostname=hostname,
subnet=host_dict.get("subnet", "corporate"),
business_criticality="high" if host_dict.get("criticality", 0.5) >= 0.7 else "medium",
status="compromised",
))
# Processes: malicious processes per compromised host
for threat in self._task_def.get("attack_chain", []):
tid = threat.get("threat_id", "T?")
for hostname in threat.get("compromised_hosts", []):
if hostname not in graph.hosts:
continue
for proc in threat.get("malicious_processes", []):
pid = f"{hostname}:{proc}"
if pid not in graph.processes:
graph.add_process(ProcessNode(
process_id=pid,
hostname=hostname,
process_name=proc,
))
# Add part_of_chain edge
graph.add_edge(Edge(
edge_type="part_of_chain",
source_id=tid,
target_id=hostname,
))
# IOCs from attack chain
for threat in self._task_def.get("attack_chain", []):
iocs = threat.get("iocs", {}) or {}
for ioc_value in iocs.get("hashes", []):
if ioc_value not in graph.iocs:
graph.add_ioc(IOCNode(ioc_value=ioc_value, ioc_type="hash", confidence=0.85))
for ioc_value in iocs.get("ips", []):
if ioc_value not in graph.iocs:
graph.add_ioc(IOCNode(ioc_value=ioc_value, ioc_type="ip", confidence=0.85))
for ioc_value in iocs.get("domains", []):
if ioc_value not in graph.iocs:
graph.add_ioc(IOCNode(ioc_value=ioc_value, ioc_type="domain", confidence=0.85))
for c2 in threat.get("c2_servers", []):
if c2 not in graph.iocs:
graph.add_ioc(IOCNode(ioc_value=c2, ioc_type="ip", confidence=0.95))
# Alerts
for a in self._task_def.get("initial_alerts", []):
aid = a.get("alert_id")
if aid and aid not in graph.alerts:
graph.add_alert(AlertNode(
alert_id=aid,
severity=a.get("severity", "medium"),
priority_score=1.0,
source_host=a.get("source_host", ""),
))
# ===========================================================================
# step()
# ===========================================================================
def step(
self,
action, # SOCActionWrapper | RedActionWrapper
timeout_s: Optional[float] = None,
**kwargs: Any,
) -> SOCObservation:
"""Process one agent action β Blue (SOCActionWrapper) or Red (RedActionWrapper).
Turn semantics (fsp_mode=True):
β’ Blue step: execute, flip active_turn β 'red', do NOT increment step_count.
β’ Red step: execute, flip active_turn β 'blue', increment step_count.
When fsp_mode=False (default / backward-compat):
β’ Blue step auto-applies a Red PassTurn so step_count always increments,
preserving all existing test and dashboard behaviour.
Returns:
SOCObservation; includes active_turn and red_observation fields.
"""
if self._state.is_done:
return self._build_observation(reward=0.0, done=True)
if isinstance(action, RedActionWrapper):
return self._step_red(action)
return self._step_blue(action)
# ------------------------------------------------------------------
# _step_blue β execute a Blue (SOC analyst) action
# ------------------------------------------------------------------
def _step_blue(
self,
action: SOCActionWrapper,
) -> SOCObservation:
"""Execute one Blue turn."""
# Convert wrapper to typed action β gracefully handle hallucinated
# action types or wrong parameters from the LLM instead of crashing.
try:
typed_action = action.to_typed_action()
except Exception as exc:
# Return a negative reward signal so GRPO can learn from the mistake
penalty = -0.2
self._state.total_reward += penalty
self._state.timeline.append({
"step": self._state.step_count + 1,
"action_type": getattr(action, "type", "unknown"),
"target": "N/A",
"result": f"INVALID_ACTION: {exc}",
"reward": penalty,
})
self._state.step_count += 1
return self._build_observation(reward=penalty, done=False)
args = typed_action.model_dump(exclude={"metadata", "type"})
# Pre-flight validation β penalise without consuming a step
current_phase = self._get_current_phase()
validation_error = self._middleware.validate(
current_phase, typed_action.type, args, self._threat_graph
)
if validation_error:
error_type = validation_error.get("error_type", "")
if error_type == "PHASE_VIOLATION":
penalty = -0.10
elif error_type == "UNJUSTIFIED_EMERGENCY":
penalty = -0.15
else:
penalty = -0.05
self._state.total_reward += penalty
return self._build_observation(reward=penalty, done=False)
# Reset per-step extras
self._last_obs_extras = {}
# Dispatch to Blue handler
reward = 0.0
result_description = "unknown action"
if isinstance(typed_action, QueryHost):
reward, result_description = self._handle_query_host(typed_action)
elif isinstance(typed_action, IsolateSegment):
reward, result_description = self._handle_isolate_segment(typed_action)
elif isinstance(typed_action, BlockIOC):
reward, result_description = self._handle_block_ioc(typed_action)
elif isinstance(typed_action, RunForensics):
reward, result_description = self._handle_run_forensics(typed_action)
elif isinstance(typed_action, KillProcess):
reward, result_description = self._handle_kill_process(typed_action)
elif isinstance(typed_action, SubmitContainmentPlan):
reward, result_description = self._handle_submit_plan(typed_action)
elif isinstance(typed_action, CorrelateAlerts):
result = self._handle_correlate_alerts(typed_action)
self._last_obs_extras.update(result)
reward = 0.05 if "error" not in result else -0.05
result_description = result.get("description", "correlate_alerts")
elif isinstance(typed_action, EnrichIOC):
result = self._handle_enrich_ioc(typed_action)
self._last_obs_extras.update(result)
reward = 0.05 if "error" not in result else -0.05
result_description = result.get("description", "enrich_ioc")
elif isinstance(typed_action, ScanHostVulnerabilities):
result = self._handle_scan_vulnerabilities(typed_action)
self._last_obs_extras.update(result)
reward = 0.05 if "error" not in result else -0.05
result_description = result.get("description", "scan_host_vulnerabilities")
elif isinstance(typed_action, TerminatePID):
reward, result_description = self._handle_terminate_pid(typed_action)
elif isinstance(typed_action, CreateFirewallRule):
reward, result_description = self._handle_create_firewall_rule(typed_action)
elif isinstance(typed_action, QuarantineFile):
reward, result_description = self._handle_quarantine_file(typed_action)
# Idempotent step reward
target = self._get_action_target(typed_action)
step_r = self._get_step_reward(
phase="investigation", action_type=typed_action.type, target=target
)
reward += step_r
self._step_reward_total += step_r
# Stall detection: penalise 3+ consecutive identical actions
stall_key = (typed_action.type, target)
if not hasattr(self, "_recent_actions"):
self._recent_actions = []
self._recent_actions.append(stall_key)
if len(self._recent_actions) >= 3:
last_three = self._recent_actions[-3:]
if last_three[0] == last_three[1] == last_three[2]:
reward -= 0.05
# Business impact grows each step (attacker progresses)
if not self._state.is_done:
impact_rate = self._task_def.get("impact_per_step", 0.02)
active_ratio = len(self._state.active_threats) / max(
1, len(self._task_def.get("attack_chain", []))
)
self._state.business_impact = min(
1.0, self._state.business_impact + impact_rate * active_ratio
)
# Round label: step_count+1 = current round being played (not yet closed)
round_label = self._state.step_count + 1
# Record timeline
self._state.timeline.append({
"step": round_label,
"action_type": typed_action.type,
"target": target,
"result": result_description,
"reward": reward,
})
# Accumulate reward
self._state.total_reward += reward
# Check if episode ends due to Blue action (plan submission)
done = False
if self._state.submitted_plan:
done = True
self._state.is_done = True
self._state.active_turn = "blue" # episode over β keep at blue
# In non-FSP mode, still increment step_count for consistency
if not self._fsp_mode:
self._state.step_count += 1
return self._build_observation(reward=reward, done=done)
# Flip turn to Red
self._state.active_turn = "red"
# fsp_mode=False (backward compat): auto-apply Red PassTurn so
# callers that only drive Blue see step_count increment as before.
if not self._fsp_mode:
# Embedded Red dynamics: execute neural or deterministic policy.
# Only fires when a policy is wired (training) or adaptive=True (SFT).
if self._neural_red_policy is not None or self._adaptive:
self._apply_red_team_dynamics(typed_action.type, target)
self._state.step_count += 1
self._state.active_turn = "blue"
# Timeout check (done after Red's "auto turn")
if self._state.step_count >= self._state.max_steps:
reward -= 0.20
self._state.total_reward -= 0.20
self._state.is_done = True
done = True
return self._build_observation(reward=reward, done=done)
# ------------------------------------------------------------------
# _step_red β execute a Red Team action
# ------------------------------------------------------------------
def _step_red(self, action: RedActionWrapper) -> SOCObservation:
"""Execute one Red turn. Only valid when active_turn == 'red'."""
if self._state.active_turn != "red":
# Wrong turn β return current obs with 0 reward (no state change)
return self._build_observation(reward=0.0, done=False)
typed_action = action.to_typed_action()
self._last_obs_extras = {}
reward = 0.0
result_description = "red: noop"
if isinstance(typed_action, LateralPivot):
reward, result_description = self._handle_lateral_pivot(typed_action)
elif isinstance(typed_action, DeployPayload):
reward, result_description = self._handle_deploy_payload(typed_action)
elif isinstance(typed_action, EvadeDetection):
reward, result_description = self._handle_evade_detection(typed_action)
elif isinstance(typed_action, PassTurn):
reward, result_description = self._handle_pass_turn(typed_action)
# Close the round: increment step_count, flip turn back to Blue
self._state.step_count += 1
self._state.active_turn = "blue"
# Record Red's action in timeline (prefixed with "red:" to distinguish)
self._state.timeline.append({
"step": self._state.step_count,
"action_type": f"red:{typed_action.type}",
"target": self._get_red_action_target(typed_action),
"result": result_description,
"reward": 0.0, # Red actions don't add to Blue's reward total
})
# Timeout check after the full round
done = False
if self._state.step_count >= self._state.max_steps:
done = True
self._state.is_done = True
return self._build_observation(reward=reward, done=done)
# ===========================================================================
# Action Handlers (return (reward, description))
# ===========================================================================
def _handle_query_host(self, action: QueryHost) -> tuple[float, str]:
"""Query a host for status info."""
hostname = action.hostname
self._last_forensics = None # Clear forensics from previous step
if hostname not in self._host_index:
return -0.05, f"Host '{hostname}' not found in network"
host = self._host_index[hostname]
# Reward for querying compromised hosts (useful investigation)
reward = 0.0
if host["status"] == "compromised" and hostname not in self._state.queried_hosts:
reward = 0.05 # Good: investigating a compromised host
elif hostname in self._state.queried_hosts:
reward = -0.02 # Penalty: re-querying same host wastes time
self._state.queried_hosts.append(hostname)
# Enhanced observation extras: process_tree + network_connections from graph
process_tree = []
if self._threat_graph is not None:
for p in self._threat_graph.processes.values():
if p.hostname == hostname:
process_tree.append({
"process_id": p.process_id,
"process_name": p.process_name,
"killed": p.killed,
})
network_connections = []
if self._threat_graph is not None:
for e in self._threat_graph.edges:
if e.edge_type == "communicates_with" and (
e.source_id == hostname or e.target_id == hostname
):
other = e.target_id if e.source_id == hostname else e.source_id
if other in self._threat_graph.iocs:
network_connections.append(other)
self._last_obs_extras["process_tree"] = process_tree
self._last_obs_extras["network_connections"] = network_connections
return reward, f"Queried {hostname}: status={host['status']}, procs={len(host['running_processes'])}"
def _handle_isolate_segment(self, action: IsolateSegment) -> tuple[float, str]:
"""Isolate a network segment, or a single host if target_host is set."""
self._last_forensics = None
# Single-host isolation path
target_host = getattr(action, "target_host", None)
if target_host:
if target_host not in self._host_index:
return -0.05, f"Host '{target_host}' not found"
self._host_index[target_host]["status"] = "isolated"
if self._threat_graph is not None and target_host in self._threat_graph.hosts:
self._threat_graph.hosts[target_host].status = "isolated"
if target_host in self._pending_followup:
self._pending_followup[target_host] = True
# Penalise isolating a clean (non-compromised) host β business disruption
compromised_host_set = {
h for threat in self._task_def["attack_chain"]
for h in threat.get("compromised_hosts", [])
}
if target_host not in compromised_host_set:
self._disruption_cost += 0.35
self._state.business_impact = min(1.0, self._state.business_impact + 0.10)
return -0.35, (
f"Isolated clean host '{target_host}' β unjustified business disruption "
f"(cumulative cost={self._disruption_cost:.2f})"
)
return 0.10, f"Isolated single host '{target_host}'"
subnet = action.subnet
if subnet not in self._network:
return -0.05, f"Subnet '{subnet}' does not exist"
if subnet in self._state.isolated_subnets:
return -0.02, f"Subnet '{subnet}' is already isolated"
# Build compromised host set for disruption tracking
compromised_host_set = {
h for threat in self._task_def["attack_chain"]
for h in threat.get("compromised_hosts", [])
}
# Isolate all hosts in the subnet; count clean hosts for disruption cost
clean_isolated_count = 0
for host in self._network[subnet]:
host["status"] = "isolated"
if self._threat_graph is not None and host["hostname"] in self._threat_graph.hosts:
self._threat_graph.hosts[host["hostname"]].status = "isolated"
if host["hostname"] in self._pending_followup:
self._pending_followup[host["hostname"]] = True
if host["hostname"] not in compromised_host_set:
clean_isolated_count += 1
self._state.isolated_subnets.append(subnet)
# Accumulate disruption cost for each clean host swept up in the isolation
if clean_isolated_count > 0:
self._disruption_cost += 0.25 * clean_isolated_count
self._state.business_impact = min(
1.0, self._state.business_impact + 0.05 * clean_isolated_count
)
# Check if this contains any active threats
reward = 0.0
threats_contained = []
for threat in self._task_def["attack_chain"]:
if threat["threat_id"] in self._state.active_threats:
# Check if any compromised hosts are in this subnet
for ch in threat["compromised_hosts"]:
if ch in self._host_index and self._host_index[ch]["subnet"] == subnet:
threats_contained.append(threat["threat_id"])
break
if threats_contained:
# Reduced reward β isolation is a blunt instrument; prefer kill_process / block_ioc
reward = 0.07 * len(threats_contained)
for tid in threats_contained:
if tid not in self._state.contained_threats:
self._state.contained_threats.append(tid)
if tid in self._state.active_threats:
self._state.active_threats.remove(tid)
# Heavy per-clean-host penalty to deter blunt-force isolation spam
if clean_isolated_count > 0:
reward -= 0.25 * clean_isolated_count
# Additional penalty for explicitly prohibited isolation
must_not_isolate = self._task_def["containment_requirements"].get("must_not_isolate", [])
if subnet in must_not_isolate:
reward -= 0.10
self._state.business_impact = min(1.0, self._state.business_impact + 0.08)
return reward, (
f"Isolated subnet '{subnet}'. Threats contained: {threats_contained}. "
f"Clean hosts disrupted: {clean_isolated_count} "
f"(cumulative cost={self._disruption_cost:.2f})"
)
def _handle_block_ioc(self, action: BlockIOC) -> tuple[float, str]:
"""Block an IOC at the perimeter.
Requires prior discovery via run_forensics or enrich_ioc; blind blocks
are recorded but yield 0 reward to prevent reward hacking.
"""
ioc = action.ioc_value
self._last_forensics = None
if ioc in self._state.blocked_iocs:
return -0.02, f"IOC '{ioc}' is already blocked"
# Prerequisite gate: IOC must have been discovered via run_forensics or enrich_ioc
if ioc not in self._discovered_iocs:
self._state.blocked_iocs.append(ioc) # record the block, but no reward
return 0.0, (
f"IOC '{ioc}' blocked without prior investigation β 0 reward "
"(run_forensics or enrich_ioc required to unlock reward)"
)
self._state.blocked_iocs.append(ioc)
# Mark forensics-confirmed hosts as responded-to β only valid for discovered IOCs,
# ensuring _pending_followup accurately reflects investigated-then-actioned flow
for hostname, responded in list(self._pending_followup.items()):
if responded:
continue
for threat in self._task_def["attack_chain"]:
if hostname in threat["compromised_hosts"]:
all_threat_iocs = (
threat["iocs"].get("hashes", [])
+ threat["iocs"].get("ips", [])
+ threat["iocs"].get("domains", [])
+ threat.get("c2_servers", [])
)
if ioc in all_threat_iocs:
self._pending_followup[hostname] = True
break
# Boosted rewards: surgical strikes are heavily preferred over blunt isolation
reward = 0.0
relevant = False
for threat in self._task_def["attack_chain"]:
all_iocs = (
threat["iocs"].get("hashes", [])
+ threat["iocs"].get("ips", [])
+ threat["iocs"].get("domains", [])
)
if ioc in all_iocs:
relevant = True
if ioc in threat.get("c2_servers", []):
reward += 0.30 # High value: severing C2 command channel
else:
reward += 0.20 # Good: blocking an investigated IOC
break
if not relevant:
reward = -0.03 # Noise: blocking irrelevant IOC
return reward, f"Blocked IOC '{ioc}' (type={action.ioc_type}). Relevant: {relevant}"
def _handle_run_forensics(self, action: RunForensics) -> tuple[float, str]:
"""Run forensic analysis on a host."""
hostname = action.hostname
if hostname not in self._host_index:
self._last_forensics = None
return -0.05, f"Host '{hostname}' not found"
host = self._host_index[hostname]
# Build forensics result based on actual host state
is_compromised = host["status"] == "compromised"
malicious_procs = []
suspicious_files = []
network_conns = []
registry_mods = []
memory_artifacts = []
if is_compromised:
# Find which threat(s) affect this host
for threat in self._task_def["attack_chain"]:
if hostname in threat["compromised_hosts"]:
malicious_procs.extend(threat["malicious_processes"])
# Generate deterministic forensic artifacts
for proc in threat["malicious_processes"]:
suspicious_files.append(f"C:\\Windows\\Temp\\{proc}.dat")
registry_mods.append(f"HKLM\\Software\\Microsoft\\Windows\\CurrentVersion\\Run\\{proc}")
for c2 in threat.get("c2_servers", []):
network_conns.append(f"{c2}:443")
for ioc_hash in threat["iocs"].get("hashes", []):
memory_artifacts.append(f"memory_inject_{ioc_hash[:8]}")
self._last_forensics = ForensicsResult(
hostname=hostname,
malicious_processes=malicious_procs,
suspicious_files=suspicious_files,
network_connections=network_conns,
registry_modifications=registry_mods,
memory_artifacts=memory_artifacts,
is_compromised=is_compromised,
)
# Reward
reward = 0.0
if hostname not in self._state.forensics_run:
if is_compromised:
reward = 0.10 # Good: found evidence
self._pending_followup.setdefault(hostname, False) # needs response action
# Reveal all IOCs tied to this host's threat chain so block_ioc can earn reward
for threat in self._task_def["attack_chain"]:
if hostname in threat.get("compromised_hosts", []):
for ioc in (
threat["iocs"].get("hashes", [])
+ threat["iocs"].get("ips", [])
+ threat["iocs"].get("domains", [])
+ threat.get("c2_servers", [])
):
self._discovered_iocs.add(ioc)
else:
reward = 0.02 # Cleared a host (some value)
self._state.forensics_run.append(hostname)
else:
reward = -0.02 # Re-running forensics wastes time
# Enhanced: behavioral_chain and network_flows from graph
behavioral_chain = []
network_flows = []
if self._threat_graph is not None:
for e in self._threat_graph.edges:
if e.source_id == hostname or e.target_id == hostname:
behavioral_chain.append({
"edge_type": e.edge_type,
"source_id": e.source_id,
"target_id": e.target_id,
})
for e in self._threat_graph.edges:
if e.edge_type == "communicates_with":
if e.source_id == hostname or e.target_id == hostname:
other = e.target_id if e.source_id == hostname else e.source_id
if other in self._threat_graph.iocs:
network_flows.append(other)
self._last_obs_extras["behavioral_chain"] = behavioral_chain
self._last_obs_extras["network_flows"] = network_flows
return reward, f"Forensics on {hostname}: compromised={is_compromised}, procs={malicious_procs}"
def _handle_kill_process(self, action: KillProcess) -> tuple[float, str]:
"""Kill a process on a host."""
hostname = action.hostname
process = action.process_name
self._last_forensics = None
if hostname not in self._host_index:
return -0.05, f"Host '{hostname}' not found"
host = self._host_index[hostname]
if host["status"] == "isolated":
return -0.02, f"Host '{hostname}' is isolated β cannot interact"
if process not in host["running_processes"]:
return -0.03, f"Process '{process}' not running on {hostname}"
# Kill the process
host["running_processes"].remove(process)
self._state.killed_processes.append({"hostname": hostname, "process": process})
if hostname in self._pending_followup:
self._pending_followup[hostname] = True
# Check if this was a malicious process
reward = 0.0
was_malicious = False
for threat in self._task_def["attack_chain"]:
if hostname in threat["compromised_hosts"] and process in threat["malicious_processes"]:
was_malicious = True
reward = 0.25 # Surgical strike: high reward for targeted process kill
# Check if all processes for this threat are killed
all_killed = True
for th_host in threat["compromised_hosts"]:
for th_proc in threat["malicious_processes"]:
still_running = (
th_host in self._host_index
and th_proc in self._host_index[th_host]["running_processes"]
)
if still_running:
all_killed = False
break
if all_killed and threat["threat_id"] in self._state.active_threats:
self._state.active_threats.remove(threat["threat_id"])
if threat["threat_id"] not in self._state.contained_threats:
self._state.contained_threats.append(threat["threat_id"])
reward += 0.15 # Bonus: fully contained a threat via surgical action
break
if not was_malicious:
reward = -0.08 # Penalty: killing legitimate process = downtime
self._state.business_impact = min(1.0, self._state.business_impact + 0.03)
return reward, f"Killed '{process}' on {hostname}. Malicious: {was_malicious}"
def _handle_terminate_pid(self, action: TerminatePID) -> tuple[float, str]:
"""Terminate a process by PID. PID is mapped to process name in this simulation."""
hostname = action.hostname
pid = action.pid
self._last_forensics = None
if hostname not in self._host_index:
return -0.05, f"Host '{hostname}' not found"
host = self._host_index[hostname]
if host["status"] == "isolated":
return -0.02, f"Host '{hostname}' is isolated - cannot interact"
process_name = pid
if ":" in pid:
pid_host, _, pid_proc = pid.partition(":")
if pid_host == hostname and pid_proc:
process_name = pid_proc
if process_name not in host["running_processes"]:
return -0.03, f"PID '{pid}' is not running on {hostname}"
host["running_processes"].remove(process_name)
self._state.killed_processes.append({"hostname": hostname, "process": process_name, "pid": pid})
if hostname in self._pending_followup:
self._pending_followup[hostname] = True
was_malicious = False
reward = 0.0
for threat in self._task_def["attack_chain"]:
if hostname in threat["compromised_hosts"] and process_name in threat["malicious_processes"]:
was_malicious = True
reward = 0.24
all_killed = True
for th_host in threat["compromised_hosts"]:
for th_proc in threat["malicious_processes"]:
if th_host in self._host_index and th_proc in self._host_index[th_host]["running_processes"]:
all_killed = False
break
if all_killed and threat["threat_id"] in self._state.active_threats:
self._state.active_threats.remove(threat["threat_id"])
if threat["threat_id"] not in self._state.contained_threats:
self._state.contained_threats.append(threat["threat_id"])
reward += 0.12
break
if not was_malicious:
reward = -0.10
self._state.business_impact = min(1.0, self._state.business_impact + 0.04)
return reward, f"Terminated benign PID '{pid}' on {hostname} - business disruption"
return reward, f"Terminated PID '{pid}' on {hostname}. Malicious: True"
def _handle_create_firewall_rule(self, action: CreateFirewallRule) -> tuple[float, str]:
"""Create firewall rule; drop blocks target IP as IOC, allow is neutral."""
hostname = action.hostname
target_ip = action.target_ip
if hostname not in self._host_index:
return -0.05, f"Host '{hostname}' not found"
if action.action == "drop":
if target_ip in self._state.blocked_iocs:
return -0.01, f"Firewall drop rule already exists for {target_ip}"
self._state.blocked_iocs.append(target_ip)
return 0.08, f"Created firewall DROP rule on {hostname} for {target_ip}"
return 0.0, f"Created firewall ALLOW rule on {hostname} for {target_ip}"
def _handle_quarantine_file(self, action: QuarantineFile) -> tuple[float, str]:
"""Quarantine suspicious files; requires terminating associated malicious PID first."""
hostname = action.hostname
file_path = action.file_path
if hostname not in self._host_index:
return -0.05, f"Host '{hostname}' not found"
file_key = (hostname, file_path)
if file_key in self._quarantined_files:
return -0.01, f"File '{file_path}' already quarantined on {hostname}"
associated_processes: List[str] = []
lowered = file_path.lower()
for threat in self._task_def.get("attack_chain", []):
if hostname not in threat.get("compromised_hosts", []):
continue
for proc in threat.get("malicious_processes", []):
expected_suffix = f"\\{proc}.dat".lower()
if lowered.endswith(expected_suffix):
associated_processes.append(proc)
if not associated_processes:
self._quarantined_files.add(file_key)
return -0.02, f"Quarantined untracked file '{file_path}' on {hostname}"
host = self._host_index[hostname]
locked = any(proc in host["running_processes"] for proc in associated_processes)
if locked:
self._state.business_impact = min(1.0, self._state.business_impact + 0.01)
return -0.04, (
f"Quarantine failed: file '{file_path}' is locked. "
"Terminate associated PID first."
)
self._quarantined_files.add(file_key)
return 0.10, f"Quarantined file '{file_path}' on {hostname}"
def _handle_submit_plan(self, action: SubmitContainmentPlan) -> tuple[float, str]:
"""Submit the final containment plan."""
self._last_forensics = None
self._state.submitted_plan = True
self._plan_entries = [entry.model_dump() for entry in action.plan]
# Grade the episode using new 10-dim grader
final_plan_dict = {
"entries": self._plan_entries,
"primary_threat_id": (self._plan_entries[0]["threat_id"]
if self._plan_entries else ""),
}
grade_result = grade_episode(
episode_actions=list(self._state.timeline),
final_plan=final_plan_dict,
graph=self._threat_graph,
task_def=self._task_def,
state=self._state,
disruption_cost=self._disruption_cost,
)
final_score = grade_result["final_score"]
# Reward proportional to final grade
reward = final_score * 1.0 # Scale: perfect score = 1.0 reward
description = (
f"Containment plan submitted. "
f"Grade: {final_score:.3f}. "
f"Threats contained: {len(self._state.contained_threats)}/{len(self._task_def['attack_chain'])}. "
f"Business impact: {self._state.business_impact:.2f}"
)
return reward, description
# ===========================================================================
# New Action Handlers (return observation-update dict)
# ===========================================================================
def _handle_correlate_alerts(self, action: CorrelateAlerts) -> dict:
"""Correlate alerts to find shared hosts/IOCs."""
if len(action.alert_ids) < 2:
return {"error": "correlate_alerts requires at least 2 alert IDs",
"description": "correlate_alerts error"}
graph = self._threat_graph
known_alerts = {aid: graph.alerts[aid] for aid in action.alert_ids if aid in graph.alerts}
if len(known_alerts) < 2:
return {"error": "fewer than 2 alert IDs found in graph",
"description": "correlate_alerts error"}
# Find shared source hosts
source_hosts: dict[str, list[str]] = {}
for aid, alert in known_alerts.items():
source_hosts.setdefault(alert.source_host, []).append(aid)
shared_hosts = [h for h, aids in source_hosts.items() if len(aids) >= 2]
# Find shared IOCs via "involves" edges
shared_iocs: set[str] = set()
for e in graph.edges:
if e.edge_type == "involves" and e.source_id in known_alerts:
if any(
e2.edge_type == "involves" and e2.target_id == e.target_id
and e2.source_id in known_alerts and e2.source_id != e.source_id
for e2 in graph.edges
):
shared_iocs.add(e.target_id)
# Update correlated_with on each alert
all_ids = list(known_alerts.keys())
for aid, alert in known_alerts.items():
for other_id in all_ids:
if other_id != aid and other_id not in alert.correlated_with:
alert.correlated_with.append(other_id)
self._state.correlated_alert_pairs.append(tuple(all_ids))
shared_count = len(shared_hosts) + len(shared_iocs)
correlation_score = min(1.0, shared_count / len(all_ids))
result = {
"correlation_results": {
"shared_hosts": shared_hosts,
"shared_iocs": list(shared_iocs),
"correlation_score": correlation_score,
},
"description": f"Correlated {len(all_ids)} alerts: {len(shared_hosts)} shared hosts",
}
return result
def _handle_enrich_ioc(self, action: EnrichIOC) -> dict:
"""Enrich an IOC with threat-intel data."""
graph = self._threat_graph
if action.ioc_value not in graph.iocs:
return {"error": "IOC not yet discovered",
"description": "enrich_ioc error"}
intel = self._task_def.get("threat_intel_data", {}) or {}
data = intel.get(action.ioc_value, {
"reputation": 0.5,
"threat_actor": "unknown",
"mitre_ttps": [],
})
# Update IOC node in graph
ioc_node = graph.iocs[action.ioc_value]
ioc_node.enriched = True
ioc_node.threat_actor = data.get("threat_actor")
ioc_node.mitre_ttps = data.get("mitre_ttps", [])
if action.ioc_value not in self._state.enriched_iocs:
self._state.enriched_iocs.append(action.ioc_value)
# Mark IOC as discovered β future block_ioc on it will receive full reward
self._discovered_iocs.add(action.ioc_value)
return {
"ioc_enrichment": data,
"description": f"Enriched IOC {action.ioc_value}: actor={data.get('threat_actor')}",
}
def _handle_scan_vulnerabilities(self, action: ScanHostVulnerabilities) -> dict:
"""Scan a host for CVE vulnerabilities."""
graph = self._threat_graph
hostname = action.hostname
if hostname not in graph.hosts:
return {"error": f"Host '{hostname}' not in Threat Graph",
"description": "scan_host_vulnerabilities error"}
vuln_chain = self._task_def.get("vulnerability_chain", []) or []
vuln_results: list[dict] = []
for entry in vuln_chain:
if not isinstance(entry, dict):
continue
if entry.get("hostname") == hostname or entry.get("affected_hosts") and hostname in entry["affected_hosts"]:
cve_id = entry.get("cve_id", "CVE-UNKNOWN")
vuln_node = VulnerabilityNode(
cve_id=cve_id,
hostname=hostname,
cvss_score=entry.get("cvss_score", 5.0),
exploitability=entry.get("exploitability", "theoretical"),
patch_available=entry.get("patch_available", False),
exploited_by_threat=entry.get("threat_id"),
)
graph.add_vulnerability(vuln_node)
graph.add_edge(Edge(
edge_type="exploits",
source_id=cve_id,
target_id=hostname,
))
vuln_results.append(entry)
# Mark host as scanned
graph.hosts[hostname].scanned = True
if hostname not in self._state.scanned_hosts:
self._state.scanned_hosts.append(hostname)
return {
"vulnerability_results": vuln_results,
"description": f"Scanned {hostname}: found {len(vuln_results)} CVEs",
}
# ===========================================================================
# Red Team Action Handlers
# ===========================================================================
def _handle_lateral_pivot(self, action: LateralPivot) -> tuple[float, str]:
"""Red: spread from a compromised host to a new target."""
src = action.source_host
dst = action.target_host
if src not in self._host_index:
return 0.0, f"red: lateral_pivot β source '{src}' not in network"
if self._host_index[src].get("status") != "compromised":
return 0.0, f"red: lateral_pivot β '{src}' not under Red control"
if dst not in self._host_index:
return 0.0, f"red: lateral_pivot β target '{dst}' not in network"
dst_status = self._host_index[dst].get("status", "online")
if dst_status == "isolated":
return 0.0, f"red: lateral_pivot β '{dst}' is isolated, pivot blocked by Blue"
if dst_status == "compromised":
return 0.0, f"red: lateral_pivot β '{dst}' already compromised"
# Compromise target and copy a process from source
self._host_index[dst]["status"] = "compromised"
src_procs = (
[p for p in self._threat_graph.processes.values() if p.hostname == src]
if self._threat_graph else []
)
proc_name = src_procs[0].process_name if src_procs else "cmd.exe"
self._host_index[dst].setdefault("running_processes", [])
if proc_name not in self._host_index[dst]["running_processes"]:
self._host_index[dst]["running_processes"].append(proc_name)
# Update threat graph
if self._threat_graph is not None:
if dst not in self._threat_graph.hosts:
hd = self._host_index[dst]
self._threat_graph.add_host(HostNode(
hostname=dst,
subnet=hd.get("subnet", "corporate"),
business_criticality="medium",
status="compromised",
))
else:
self._threat_graph.hosts[dst].status = "compromised"
pid = f"{dst}:{proc_name}"
if pid not in self._threat_graph.processes:
self._threat_graph.add_process(ProcessNode(
process_id=pid, hostname=dst, process_name=proc_name
))
self._threat_graph.add_edge(Edge(
edge_type="pivoted_from", source_id=dst, target_id=src
))
# Generate SIEM alert for Blue
alert_id = f"PIVOT-{uuid.uuid4().hex[:6].upper()}"
subnet = self._host_index.get(dst, {}).get("subnet", "unknown")
self._alert_queue.append({
"alert_id": alert_id,
"timestamp": "2024-01-01T00:00:00Z",
"source_host": dst,
"severity": "critical",
"threat_type": "lateral_movement",
"description": (
f"Lateral movement detected: {proc_name} spawned on {dst} "
f"(pivot from {src})"
),
"ioc_indicators": [],
"subnet": subnet,
"is_acknowledged": False,
})
if self._threat_graph is not None:
self._threat_graph.add_alert(AlertNode(
alert_id=alert_id, severity="critical",
priority_score=15.0, source_host=dst,
))
# Update live rubric
if self._live_requirements is not None:
self._live_requirements.setdefault("must_kill", []).append({
"hostname": dst, "process": proc_name, "threat_id": "FSP_PIVOT",
})
return 0.0, f"red: lateral_pivot {src} β {dst} (proc={proc_name})"
def _handle_deploy_payload(self, action: DeployPayload) -> tuple[float, str]:
"""Red: deploy a malicious payload on a host Red controls."""
hostname = action.hostname
payload_type = action.payload_type
if hostname not in self._host_index:
return 0.0, f"red: deploy_payload β '{hostname}' not in network"
if self._host_index[hostname].get("status") != "compromised":
return 0.0, f"red: deploy_payload β no shell on '{hostname}'"
proc_name = {
"ransomware": "ransomware.exe",
"exfiltration": "exfil_agent.exe",
"c2": "c2_beacon.exe",
}[payload_type]
host = self._host_index[hostname]
if proc_name not in host.get("running_processes", []):
host.setdefault("running_processes", []).append(proc_name)
if self._threat_graph is not None:
pid = f"{hostname}:{proc_name}"
if pid not in self._threat_graph.processes:
self._threat_graph.add_process(ProcessNode(
process_id=pid, hostname=hostname, process_name=proc_name
))
impact_delta = {"ransomware": 0.15, "exfiltration": 0.08, "c2": 0.05}[payload_type]
self._state.business_impact = min(1.0, self._state.business_impact + impact_delta)
severity = {"ransomware": "critical", "exfiltration": "high", "c2": "high"}[payload_type]
alert_id = f"PAYLOAD-{uuid.uuid4().hex[:6].upper()}"
self._alert_queue.append({
"alert_id": alert_id,
"timestamp": "2024-01-01T00:00:00Z",
"source_host": hostname,
"severity": severity,
"threat_type": payload_type,
"description": (
f"{payload_type.capitalize()} payload deployed on {hostname}: {proc_name}"
),
"ioc_indicators": [],
"subnet": host.get("subnet", "unknown"),
"is_acknowledged": False,
})
if self._threat_graph is not None:
self._threat_graph.add_alert(AlertNode(
alert_id=alert_id, severity=severity,
priority_score=18.0, source_host=hostname,
))
return 0.0, f"red: deployed {payload_type} payload on {hostname}"
def _handle_evade_detection(self, action: EvadeDetection) -> tuple[float, str]:
"""Red: apply a detection-evasion technique on a controlled host."""
hostname = action.hostname
technique = action.technique
if hostname not in self._host_index:
return 0.0, f"red: evade_detection β '{hostname}' not in network"
if self._host_index[hostname].get("status") != "compromised":
return 0.0, f"red: evade_detection β no shell on '{hostname}'"
if technique == "migrate_pid":
host = self._host_index[hostname]
malicious_procs = {
proc
for threat in self._task_def.get("attack_chain", [])
if hostname in threat.get("compromised_hosts", [])
for proc in threat.get("malicious_processes", [])
}
for i, proc in enumerate(list(host.get("running_processes", []))):
if proc in malicious_procs:
new_name = f"svchost_{i}.exe"
host["running_processes"][i] = new_name
if self._threat_graph:
old_pid = f"{hostname}:{proc}"
if old_pid in self._threat_graph.processes:
self._threat_graph.processes.pop(old_pid)
new_pid = f"{hostname}:{new_name}"
self._threat_graph.add_process(ProcessNode(
process_id=new_pid, hostname=hostname,
process_name=new_name,
))
return 0.0, f"red: migrated PIDs on {hostname} to blend with system processes"
if technique == "clear_logs":
before = len(self._alert_queue)
self._alert_queue = [
a for a in self._alert_queue
if a.get("source_host") != hostname
]
removed = before - len(self._alert_queue)
return 0.0, f"red: cleared {removed} SIEM alert(s) from {hostname}"
return 0.0, f"red: evasion '{technique}' applied on {hostname}"
def _handle_pass_turn(self, action: PassTurn) -> tuple[float, str]: # noqa: ARG002
"""Red: remain stealthy, take no action."""
return 0.0, "red: pass_turn (stealth)"
def _get_red_action_target(self, action: Any) -> str:
"""Extract a compact target string from a Red action for timeline logging."""
if isinstance(action, LateralPivot):
return f"{action.source_host}β{action.target_host}"
if isinstance(action, DeployPayload):
return f"{action.hostname}/{action.payload_type}"
if isinstance(action, EvadeDetection):
return f"{action.hostname}/{action.technique}"
return "β"
# ===========================================================================
# Helpers
# ===========================================================================
def _compute_reward_dimensions(self) -> Dict[str, float]:
"""Per-step heuristic partial scores for all 10 grading dimensions.
Evidence-gated: actions only score if prior evidence justified them.
Result-usage: forensics-confirmed hosts with no followup are penalized.
Scores in [0, 1]; terminal grade_breakdown supersedes these on plan submission.
"""
state = self._state
task_chain = self._task_def.get("attack_chain", [])
total_threats = max(1, len(task_chain))
total_compromised = max(1, sum(len(t.get("compromised_hosts", [])) for t in task_chain))
total_iocs = max(1, sum(
len(t.get("iocs", {}).get("hashes", []))
+ len(t.get("iocs", {}).get("ips", []))
+ len(t.get("iocs", {}).get("domains", []))
for t in task_chain
))
# --- Build evidence pools: what the agent could have observed ---
# Hosts mentioned as alert source (visible from turn 0)
alert_source_hosts: set = set()
for a in self._task_def.get("initial_alerts", []):
alert_source_hosts.add(a.get("source_host", ""))
for a in self._alert_queue:
alert_source_hosts.add(a.get("source_host", ""))
alert_source_hosts.discard("")
# IOCs visible from alert ioc_indicators
alert_iocs: set = set()
for a_list in (self._task_def.get("initial_alerts", []), self._alert_queue):
for a in a_list:
for ioc in a.get("ioc_indicators", []):
alert_iocs.add(ioc)
# IOCs revealed by running forensics on a host
forensics_revealed_iocs: set = set()
for hostname in state.forensics_run:
for threat in task_chain:
if hostname in threat.get("compromised_hosts", []):
forensics_revealed_iocs.update(threat.get("c2_servers", []))
forensics_revealed_iocs.update(threat["iocs"].get("hashes", []))
forensics_revealed_iocs.update(threat["iocs"].get("ips", []))
forensics_revealed_iocs.update(threat["iocs"].get("domains", []))
discovered_iocs = alert_iocs | forensics_revealed_iocs
# 1. threat_containment β fraction of threats neutralised (no evidence gate; outcome IS evidence)
threat_containment = min(1.0, len(state.contained_threats) / total_threats)
# 2. ioc_blocking β only blocks of IOCs the agent actually discovered count
justified_blocks = [ioc for ioc in state.blocked_iocs if ioc in discovered_iocs]
ioc_blocking = min(1.0, len(justified_blocks) / total_iocs)
# 3. forensic_investigation β only counts forensics on alert-mentioned or previously queried
# hosts; penalizes confirmed compromises left with no response action
justified_forensics = [
h for h in state.forensics_run
if h in alert_source_hosts or h in state.queried_hosts
]
pending = self._pending_followup
unresponded = sum(1 for v in pending.values() if not v)
followup_penalty = min(0.30, unresponded * 0.10)
forensic_investigation = max(0.0,
min(1.0, len(justified_forensics) / total_compromised) - followup_penalty
)
# 4. siem_correlation β scored by semantic quality (shared source hosts or IOCs)
if not state.correlated_alert_pairs:
siem_correlation = 0.0
else:
alert_map: Dict[str, Any] = {}
for a in self._task_def.get("initial_alerts", []):
alert_map[a.get("alert_id", "")] = a
for a in self._alert_queue:
alert_map[a.get("alert_id", "")] = a
quality_scores = []
for pair in state.correlated_alert_pairs:
pair_alerts = [alert_map[aid] for aid in pair if aid in alert_map]
if len(pair_alerts) < 2:
quality_scores.append(0.3)
continue
sources = [a.get("source_host") for a in pair_alerts]
ioc_sets = [set(a.get("ioc_indicators", [])) for a in pair_alerts]
shared_hosts = len(sources) != len({s for s in sources if s})
shared_iocs = bool(ioc_sets[0] & ioc_sets[1]) if len(ioc_sets) >= 2 else False
quality_scores.append(1.0 if (shared_hosts or shared_iocs) else 0.2)
siem_correlation = sum(quality_scores) / max(1, len(quality_scores))
# 5. threat_intel_usage β only enrichments of discovered IOCs count
justified_enrichments = [ioc for ioc in state.enriched_iocs if ioc in discovered_iocs]
threat_intel_usage = min(1.0, len(justified_enrichments) / total_iocs)
# 6. vuln_root_cause β fraction of threats with a scanned host
vuln_root_cause = min(1.0, len(state.scanned_hosts) / total_threats)
# 7. business_impact β proportionate isolation + low overall impact
# Reward: isolating confirmed-compromised hosts Penalize: isolating clean hosts
isolated_host_set = {
h for h, hd in self._host_index.items() if hd.get("status") == "isolated"
} if self._host_index else set()
compromised_host_set = {
h for threat in task_chain for h in threat.get("compromised_hosts", [])
}
if isolated_host_set:
over_isolated = isolated_host_set - compromised_host_set
isolation_proportion = (
len(isolated_host_set - over_isolated) / len(isolated_host_set)
)
over_iso_penalty = min(0.40, len(over_isolated) * 0.15)
else:
isolation_proportion = 1.0
over_iso_penalty = 0.0
raw_impact_score = max(0.0, 1.0 - state.business_impact)
business_impact = max(0.0, min(1.0,
0.6 * raw_impact_score + 0.4 * isolation_proportion - over_iso_penalty
))
# 8. step_efficiency β reward early resolution
ratio = state.step_count / max(1, state.max_steps)
step_efficiency = max(0.0, 1.0 - max(0.0, ratio - 0.5) * 1.5)
# 9. plan_coverage β partial credit scales with threats addressed
if state.submitted_plan:
plan_coverage = min(1.0, len(self._plan_entries) / total_threats)
else:
plan_coverage = min(0.5, len(state.contained_threats) / total_threats * 0.5)
# 10. plan_evidence_quality β confidence of submitted plan; else evidence depth proxy
if state.submitted_plan and self._plan_entries:
avg_conf = sum(e.get("confidence", 0.0) for e in self._plan_entries) / len(self._plan_entries)
plan_evidence_quality = float(avg_conf)
else:
evidence_count = len(justified_forensics) + len(justified_enrichments) + len(state.scanned_hosts)
plan_evidence_quality = min(0.5, evidence_count / (total_compromised * 3) * 0.5)
return {
"threat_containment": round(threat_containment, 4),
"ioc_blocking": round(ioc_blocking, 4),
"forensic_investigation": round(forensic_investigation, 4),
"siem_correlation": round(siem_correlation, 4),
"threat_intel_usage": round(threat_intel_usage, 4),
"vuln_root_cause": round(vuln_root_cause, 4),
"business_impact": round(business_impact, 4),
"step_efficiency": round(step_efficiency, 4),
"plan_coverage": round(plan_coverage, 4),
"plan_evidence_quality": round(plan_evidence_quality, 4),
}
def _get_current_phase(self) -> str:
"""Derive episode phase from the action history in the timeline."""
action_types = {t["action_type"] for t in self._state.timeline}
if any(t in action_types for t in ["kill_process", "block_ioc", "isolate_segment", "terminate_pid", "create_firewall_rule", "quarantine_file"]):
return "remediation"
if any(t in action_types for t in ["run_forensics", "enrich_ioc", "scan_host_vulnerabilities", "query_host"]):
return "investigation"
return "triage"
def _build_observation(self, reward: float, done: bool) -> SOCObservation:
"""Build the observation from current state."""
# Compute network topology summary
subnet_counts = {name: len(hosts) for name, hosts in self._network.items()}
compromised = sum(
1 for hosts in self._network.values()
for h in hosts if h["status"] == "compromised"
)
isolated = sum(
1 for hosts in self._network.values()
for h in hosts if h["status"] == "isolated"
)
total = sum(len(hosts) for hosts in self._network.values())
topology = NetworkTopology(
total_hosts=total,
subnets=subnet_counts,
compromised_count=compromised,
isolated_count=isolated,
online_count=total - compromised - isolated,
)
# Build alert list
alerts = [Alert(**a) for a in self._alert_queue]
# Build timeline
timeline = [
TimelineEntry(
step=t["step"],
action_type=t["action_type"],
target=t["target"],
result=t["result"],
reward=t["reward"],
)
for t in self._state.timeline
]
# Compute final grade if done
final_score_val = None
grade_breakdown_val = None
if done and self._state.submitted_plan:
final_plan_dict = {
"entries": self._plan_entries,
"primary_threat_id": (self._plan_entries[0]["threat_id"]
if self._plan_entries else ""),
}
computed = grade_episode(
episode_actions=list(self._state.timeline),
final_plan=final_plan_dict,
graph=self._threat_graph,
task_def=self._task_def,
state=self._state,
disruption_cost=self._disruption_cost,
)
final_score_val = round(computed["final_score"], 4)
grade_breakdown_val = computed["breakdown"]
# Merge per-step observation extras (process_tree, correlation_results, etc.)
extras = getattr(self, "_last_obs_extras", {}) or {}
threat_graph_summary = None
if self._threat_graph is not None:
threat_graph_summary = self._threat_graph.get_context_summary()
# Per-step partial reward dimensions for GRPO credit assignment
reward_dimensions = self._compute_reward_dimensions()
# Red observation β only populated when it is Red's turn next
red_obs = (
self._generate_red_observation()
if self._state.active_turn == "red"
else None
)
return SOCObservation(
episode_id=self._state.episode_id or "",
alert_queue=alerts,
network_topology=topology,
host_forensics=self._last_forensics,
timeline=timeline,
business_impact_score=round(self._state.business_impact, 4),
step_count=self._state.step_count,
active_threats=list(self._state.active_threats),
max_steps=self._state.max_steps,
task_id=self._state.task_id,
total_reward=round(self._state.total_reward, 4),
final_score=final_score_val,
grade_breakdown=grade_breakdown_val,
done=done,
reward=round(reward, 4),
correlation_results=extras.get("correlation_results"),
ioc_enrichment=extras.get("ioc_enrichment"),
vulnerability_results=extras.get("vulnerability_results"),
playbook_result=None,
threat_graph_summary=threat_graph_summary,
available_playbooks=[],
reward_dimensions=reward_dimensions,
active_turn=self._state.active_turn,
red_observation=red_obs,
)
def _get_action_target(self, action: Any) -> str:
"""Extract the target string from a typed action for timeline logging."""
if isinstance(action, QueryHost):
return action.hostname
elif isinstance(action, IsolateSegment):
return getattr(action, "target_host", None) or action.subnet
elif isinstance(action, BlockIOC):
return f"{action.ioc_type}:{action.ioc_value}"
elif isinstance(action, RunForensics):
return action.hostname
elif isinstance(action, KillProcess):
return f"{action.hostname}/{action.process_name}"
elif isinstance(action, SubmitContainmentPlan):
return f"{len(action.plan)} entries"
elif isinstance(action, CorrelateAlerts):
return ",".join(action.alert_ids)
elif isinstance(action, EnrichIOC):
return action.ioc_value
elif isinstance(action, ScanHostVulnerabilities):
return action.hostname
elif isinstance(action, TerminatePID):
return f"{action.hostname}/{action.pid}"
elif isinstance(action, CreateFirewallRule):
return f"{action.hostname}:{action.action}:{action.target_ip}"
elif isinstance(action, QuarantineFile):
return f"{action.hostname}:{action.file_path}"
return "unknown"
# ===========================================================================
# Adaptive Red Team + Step Rewards (Task 10)
# ===========================================================================
def _generate_red_observation(self) -> Dict[str, Any]:
"""What the Red Team LLM sees: footholds it controls + Blue's last action.
Returned as the ``red_observation`` field in SOCObservation whenever
``active_turn == 'red'``, so inference.py can feed it straight to the
Red LLM without a separate API call.
"""
compromised_hosts = [
h for h, hd in self._host_index.items()
if hd.get("status") == "compromised"
]
# Most recent Blue action from the timeline (exclude Red's own entries)
blue_actions_detected: List[Dict[str, Any]] = []
for entry in reversed(self._state.timeline):
action_type = entry.get("action_type", "")
if not action_type.startswith("red:"):
blue_actions_detected.append({
"step": entry["step"],
"action": action_type,
"target": entry["target"],
"result": entry["result"],
})
break # Only the single most recent Blue action
return {
"episode_id": self._state.episode_id,
"round": self._state.step_count + 1,
"compromised_hosts": compromised_hosts,
"blue_actions_detected": blue_actions_detected,
"active_threats": list(self._state.active_threats),
"business_impact": round(self._state.business_impact, 4),
}
def _log_red_decision(self, observation: Dict[str, Any], action: Dict[str, Any]) -> None:
"""Record (observation -> action) tuples for red-team imitation warm-start."""
record = {"observation": observation, "action": action}
self._red_team_decisions.append(record)
if self._red_team_logger is not None:
try:
self._red_team_logger(record)
except Exception:
# Logging is best effort and should never affect environment execution.
pass
def _apply_red_team_dynamics(self, action_type: str, target: str) -> None:
"""Execute embedded Red dynamics in non-FSP mode.
When neural_red_policy is callable: invoke it with the current red
observation, route the returned action through the Red handlers, and
log the (obs β action) pair for offline SFT.
When neural_red_policy is None (adaptive=True path): apply the
deterministic fallback policy and log the pair.
"""
red_obs = self._generate_red_observation()
if callable(self._neural_red_policy):
try:
action_dict = self._neural_red_policy(red_obs)
if not isinstance(action_dict, dict):
action_dict = {"type": "pass_turn"}
except Exception:
action_dict = {"type": "pass_turn"}
atype = action_dict.get("type", "pass_turn")
if atype == "lateral_pivot":
src = action_dict.get("source_host", "")
dst = action_dict.get("target_host", "")
if src and dst:
self._handle_lateral_pivot(
LateralPivot(type="lateral_pivot", source_host=src, target_host=dst)
)
elif atype == "deploy_payload":
h = action_dict.get("hostname", "")
pl = action_dict.get("payload_type", "ransomware")
if h:
self._handle_deploy_payload(
DeployPayload(type="deploy_payload", hostname=h, payload_type=pl)
)
elif atype == "evade_detection":
h = action_dict.get("hostname", "")
tech = action_dict.get("technique", "migrate_pid")
if h:
self._handle_evade_detection(
EvadeDetection(type="evade_detection", hostname=h, technique=tech)
)
# pass_turn β no graph mutation needed
self._log_red_decision(red_obs, action_dict)
else:
# Deterministic fallback for imitation warm-start (adaptive=True path)
det_action = self._deterministic_red_policy(action_type, target, red_obs)
atype = det_action.get("type", "pass_turn")
if atype == "lateral_pivot":
self._handle_lateral_pivot(
LateralPivot(
type="lateral_pivot",
source_host=det_action["source_host"],
target_host=det_action["target_host"],
)
)
elif atype == "deploy_payload":
dp_host = det_action.get("hostname", "")
dp_payload = det_action.get("payload_type", "ransomware")
if dp_host:
self._handle_deploy_payload(
DeployPayload(
type="deploy_payload",
hostname=dp_host,
payload_type=dp_payload,
)
)
self._log_red_decision(red_obs, det_action)
def _deterministic_red_policy(
self, blue_action: str, blue_target: str, red_obs: Dict[str, Any]
) -> Dict[str, Any]:
"""Rule-based Red policy for SFT imitation warm-start data collection.
Priority order:
1. Stall punishment β >= 3 consecutive passive Blue actions deploy ransomware.
2. Reactive pivot β Blue containment action triggers lateral movement.
3. Autonomous pivot β 15% chance to spread even on passive Blue actions.
"""
_passive = frozenset({"query_host", "pass_turn"})
_containment = frozenset({"kill_process", "isolate_segment", "block_ioc"})
compromised = red_obs.get("compromised_hosts", [])
# 1. Stall punishment: >= 3 consecutive passive steps without containment
if blue_action in _passive and compromised:
streak = 0
for entry in reversed(getattr(self, "_recent_actions", [])):
if isinstance(entry, tuple) and entry[0] in _passive:
streak += 1
else:
break
if streak >= 3:
return {
"type": "deploy_payload",
"hostname": compromised[0],
"payload_type": "ransomware",
}
# 2. Reactive pivot on Blue containment actions
if blue_action in _containment:
src = compromised[0] if compromised else (blue_target or None)
if src is not None and src in self._host_index:
dst = next(
(h for h, hd in self._host_index.items()
if hd.get("status") not in ("compromised", "isolated") and h != src),
None,
)
if dst:
return {"type": "lateral_pivot", "source_host": src, "target_host": dst}
# 3. Autonomous pivot: 15% chance even when Blue is passive
if blue_action in _passive and compromised and self._rng.random() < 0.15:
src = compromised[0]
dst = next(
(h for h, hd in self._host_index.items()
if hd.get("status") not in ("compromised", "isolated") and h != src),
None,
)
if dst:
return {"type": "lateral_pivot", "source_host": src, "target_host": dst}
return {"type": "pass_turn"}
def export_red_team_decisions(self) -> List[Dict[str, Any]]:
"""Return a copy of recorded red-team decisions for offline SFT."""
return list(self._red_team_decisions)
STEP_REWARDS: Dict[Any, float] = {
("investigation", "run_forensics"): +0.10,
("investigation", "enrich_ioc"): +0.05,
("investigation", "scan_host_vulnerabilities"): +0.05,
("triage", "correlate_alerts"): +0.05,
"phase_violation_attempt": -0.20,
"ungrounded_action_attempt": -0.10,
}
def _get_step_reward(self, phase: str, action_type: str, target: str) -> float:
"""Idempotent step reward β fires only once per (phase, action_type, target) triple.
Hard cap: total step rewards per episode never exceed 0.40.
"""
if not hasattr(self, "_fired_step_rewards"):
self._fired_step_rewards = set()
# Hard cap: once we've reached 0.40 in step rewards, return 0 for all subsequent
if getattr(self, "_step_reward_total", 0.0) >= 0.40:
return 0.0
key = (phase, action_type, target)
if key in self._fired_step_rewards:
return 0.0
reward = self.STEP_REWARDS.get((phase, action_type), 0.0)
if reward != 0.0:
self._fired_step_rewards.add(key)
return reward
def _maybe_reinfect(self, hostname: str, process_name: str) -> None:
"""30 % chance to reinfect with a _v2 variant when unblocked IOCs exist in the threat chain."""
if not self._adaptive:
return
graph = self._threat_graph
if graph is None:
return
# Check whether any IOC in the host's threat chain is still unblocked
unblocked_chain_iocs = False
for ioc_node in graph.iocs.values():
if not ioc_node.blocked:
# Is this IOC linked (via any edge) to the same host's chain?
for e in graph.edges:
if e.target_id == hostname or e.source_id == hostname:
unblocked_chain_iocs = True
break
if unblocked_chain_iocs:
break
if not unblocked_chain_iocs:
return
if self._rng.random() >= 0.3:
return
# Reinfect: spawn a _v2 variant process on the host
variant_name = f"{process_name}_v2"
if hostname in self._host_index:
host = self._host_index[hostname]
if variant_name not in host["running_processes"]:
host["running_processes"].append(variant_name)
host["status"] = "compromised"
# Add the variant to the threat graph
pid = f"{hostname}:{variant_name}"
if pid not in graph.processes:
graph.add_process(ProcessNode(
process_id=pid,
hostname=hostname,
process_name=variant_name,
killed=False,
))
# Emit a CRITICAL alert to signal the reinfection
alert_id = f"REINFECT-{uuid.uuid4().hex[:6].upper()}"
graph.add_alert(AlertNode(
alert_id=alert_id,
severity="critical",
priority_score=18.0,
source_host=hostname,
))
self._alert_queue.append({
"alert_id": alert_id,
"timestamp": "2024-01-01T00:00:00Z",
"source_host": hostname,
"severity": "critical",
"threat_type": "malware",
"description": f"Reinfection detected: {variant_name} spawned on {hostname} (IOC-assisted persistence)",
"ioc_indicators": [],
"subnet": self._host_index.get(hostname, {}).get("subnet", "unknown"),
"is_acknowledged": False,
})
def _adversary_react(self, action_type: str, target: str) -> Optional[Dict[str, Any]]:
"""Legacy hook β disabled; Red Team now acts via explicit RedActionWrapper steps."""
return None
@property
def state(self) -> SOCState:
"""Get the current internal environment state."""
return self._state
|