Spaces:
Sleeping
Sleeping
File size: 75,082 Bytes
115612d d73d6cb 115612d d73d6cb 115612d d73d6cb 115612d d73d6cb 115612d bb65dd1 852d63f bb65dd1 115612d | 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 | """
OpenEnv-CloudSOC: Cloud Security Operations Center Benchmark Environment
=========================================================================
A Gymnasium-compatible environment for evaluating LLM agents on cloud security
incident response. Implements Zero-DB architecture with in-memory state management.
Hardware Target: 2 vCPU / 8 GB RAM Docker container
Architecture: Thin Client, Deep State (no external databases)
Key Features:
- Realistic AWS cloud simulation with deceptive elements
- Progressive reward shaping through forensic kill-chain
- Strict causal dependencies for proper IR procedure
- Adversarial traps to test agent judgment
- Multi-task campaign with state persistence
- Comprehensive audit trail and analytics
"""
import random
import json
import hashlib
import copy
from datetime import datetime, timedelta
from typing import Dict, List, Optional, Any, Tuple, Union, Set
from enum import Enum
from dataclasses import dataclass, field
from pydantic import BaseModel, Field, validator
from collections import defaultdict
import gymnasium as gym
from gymnasium import spaces
class Observation(BaseModel):
state_description: str
metadata: Optional[Dict[str, Any]] = None
class Action(BaseModel):
tool: str
args: Dict[str, Any]
class Reward(BaseModel):
value: float
reason: Optional[str] = None
# =============================================================================
# CONSTANTS AND CONFIGURATION
# =============================================================================
# Reward constants for fine-tuned scoring
REWARDS = {
"FLAG_DISCOVERY": 0.02,
"CRITICAL_ACTION": 0.10,
"CONTAINMENT_ACTION": 0.05,
"PARSE_ERROR": -0.02,
"PRECONDITION_FAIL": -0.05,
"EVIDENCE_DESTRUCTION": -1.0,
"QUERY_BASIC": -0.01,
"QUERY_DEEP": -0.05,
"INVALID_TOOL": -0.01,
"TIMELINE_BONUS_MAX": 0.5,
}
# Malicious IPs for realistic simulation
MALICIOUS_IPS = [
"185.220.101.42", # TOR exit node
"45.155.205.233", # Known C2
"194.26.192.64", # Bulletproof hosting
]
# Benign IPs for noise
BENIGN_IPS = [
"10.0.0.1", "10.0.1.50", "10.0.5.20", "172.16.0.100",
"192.168.1.10", "10.0.0.254", "172.31.0.1"
]
# =============================================================================
# PYDANTIC MODELS FOR TOOL ABSTRACTION (Mechanic #7)
# =============================================================================
class ToolArgs(BaseModel):
"""Base class for tool arguments"""
class Config:
extra = "forbid"
class CloudWatchQueryArgs(ToolArgs):
log_group: str = Field(..., description="CloudWatch log group to query")
start_time: Optional[str] = Field(None, description="ISO timestamp for query start")
end_time: Optional[str] = Field(None, description="ISO timestamp for query end")
filter_pattern: Optional[str] = Field(None, description="CloudWatch filter pattern")
class EC2ActionArgs(ToolArgs):
instance_id: str = Field(..., pattern=r"^i-[a-f0-9]+$")
class IAMRoleArgs(ToolArgs):
role_name: str = Field(..., min_length=1, max_length=64)
class IAMDetachArgs(ToolArgs):
role_name: str
instance_id: Optional[str] = None
class S3BucketArgs(ToolArgs):
bucket_name: str = Field(..., pattern=r"^[a-z0-9][a-z0-9.-]{1,61}[a-z0-9]$")
class CredentialRotateArgs(ToolArgs):
resource_type: str = Field(..., description="Type: 'rds', 'iam', 's3'")
resource_id: str
class SnapshotArgs(ToolArgs):
instance_id: str = Field(..., pattern=r"^i-[a-f0-9]+$")
snapshot_type: str = Field(default="forensic", description="'forensic' or 'backup'")
class IncidentCloseArgs(ToolArgs):
incident_id: str
timeline: List[str] = Field(..., min_items=1, description="Ordered list of incident events")
root_cause: Optional[str] = None
remediation_summary: Optional[str] = None
class SecurityGroupArgs(ToolArgs):
security_group_id: str = Field(..., pattern=r"^sg-[a-f0-9]+$")
action: str = Field(..., description="'isolate' or 'restore'")
class InvestigateArgs(ToolArgs):
target_type: str = Field(..., description="'ip', 'user', 'role', 'instance'")
target_id: str
class ToolCall(BaseModel):
"""Strict tool call schema that agents must use"""
thought: str = Field(..., description="1-sentence internal reasoning", max_length=500)
tool: str = Field(..., description="Tool name to invoke")
args: Dict[str, Any] = Field(default_factory=dict)
@validator('tool')
def validate_tool_name(cls, v):
valid_tools = {
'aws.cloudwatch.query_basic', 'aws.cloudwatch.query_deep',
'aws.ec2.isolate', 'aws.ec2.terminate', 'aws.ec2.describe',
'aws.ec2.snapshot', 'aws.iam.detach_role', 'aws.iam.revoke_credentials',
'aws.iam.describe_role', 'aws.iam.list_policies',
'aws.s3.get_bucket_policy', 'aws.s3.block_public_access',
'aws.s3.list_objects', 'aws.rds.rotate_credentials',
'aws.security_group.modify', 'aws.investigate',
'aws.soc.close_incident', 'aws.soc.get_alerts',
'aws.guardduty.get_findings', 'aws.cloudtrail.lookup_events',
'aws.config.get_compliance', 'aws.ssm.run_command',
'aws.lambda.list_functions', 'aws.sts.get_caller_identity'
}
if v not in valid_tools:
raise ValueError(f"Unknown tool: {v}. Valid tools: {sorted(valid_tools)}")
return v
@validator('thought')
def validate_thought(cls, v):
if not v or len(v.strip()) < 5:
raise ValueError("Thought must contain meaningful reasoning (min 5 chars)")
return v.strip()
# =============================================================================
# STATE MODELS (In-Memory Cloud State)
# =============================================================================
class InstanceState(Enum):
RUNNING = "running"
STOPPED = "stopped"
ISOLATED = "isolated"
TERMINATED = "terminated"
class IncidentPhase(Enum):
INVESTIGATION = "investigation"
CONTAINMENT = "containment"
ERADICATION = "eradication"
RECOVERY = "recovery"
CLOSED = "closed"
@dataclass
class EC2Instance:
instance_id: str
state: InstanceState = InstanceState.RUNNING
attached_roles: List[str] = field(default_factory=list)
security_groups: List[str] = field(default_factory=list)
is_compromised: bool = False
has_forensic_snapshot: bool = False
metadata: Dict[str, Any] = field(default_factory=dict)
@dataclass
class IAMRole:
role_name: str
policies: List[str] = field(default_factory=list)
is_compromised: bool = False
is_detached: bool = False
credentials_rotated: bool = False
has_backdoor: bool = False
@dataclass
class S3Bucket:
bucket_name: str
is_public: bool = False
contains_credentials: bool = False
leaked_keys: List[str] = field(default_factory=list)
public_access_blocked: bool = False
@dataclass
class LogEntry:
timestamp: str
source: str
message: str
severity: str # INFO, WARN, ERROR, CRITICAL
is_attack_indicator: bool = False
is_noise: bool = False
is_red_herring: bool = False
@dataclass
class Alert:
alert_id: str
title: str
severity: str
source: str
timestamp: str
details: Dict[str, Any] = field(default_factory=dict)
is_true_positive: bool = True
@dataclass
class CloudState:
"""Complete in-memory cloud infrastructure state"""
instances: Dict[str, EC2Instance] = field(default_factory=dict)
roles: Dict[str, IAMRole] = field(default_factory=dict)
buckets: Dict[str, S3Bucket] = field(default_factory=dict)
security_groups: Dict[str, Dict] = field(default_factory=dict)
logs: List[LogEntry] = field(default_factory=list)
alerts: List[Alert] = field(default_factory=list)
credentials: Dict[str, Dict] = field(default_factory=dict)
# Investigation state
discovered_flags: set = field(default_factory=set)
completed_actions: List[str] = field(default_factory=list)
incident_phase: IncidentPhase = IncidentPhase.INVESTIGATION
# Timeline tracking for grading
agent_timeline: List[str] = field(default_factory=list)
ground_truth_timeline: List[str] = field(default_factory=list)
# Enhanced tracking
evidence_collected: Dict[str, Any] = field(default_factory=dict)
mitre_techniques: Set[str] = field(default_factory=set)
ioc_indicators: List[Dict] = field(default_factory=list)
def to_dict(self) -> Dict:
"""Serialize state for debugging/checkpointing"""
return {
"instances": {k: self._inst_to_dict(v) for k, v in self.instances.items()},
"roles": {k: self._role_to_dict(v) for k, v in self.roles.items()},
"buckets": {k: self._bucket_to_dict(v) for k, v in self.buckets.items()},
"discovered_flags": list(self.discovered_flags),
"completed_actions": self.completed_actions,
"phase": self.incident_phase.value,
"evidence_collected": self.evidence_collected,
}
def _inst_to_dict(self, inst: EC2Instance) -> Dict:
return {
"instance_id": inst.instance_id, "state": inst.state.value,
"is_compromised": inst.is_compromised, "has_snapshot": inst.has_forensic_snapshot
}
def _role_to_dict(self, role: IAMRole) -> Dict:
return {
"role_name": role.role_name, "is_compromised": role.is_compromised,
"is_detached": role.is_detached, "has_backdoor": role.has_backdoor
}
def _bucket_to_dict(self, bucket: S3Bucket) -> Dict:
return {
"bucket_name": bucket.bucket_name, "is_public": bucket.is_public,
"public_access_blocked": bucket.public_access_blocked
}
# =============================================================================
# SCENARIO DEFINITIONS (Multi-Task Campaign - Mechanic #11)
# =============================================================================
SCENARIOS = {
"easy": {
"name": "Leaky S3 Bucket Discovery",
"description": "Identify and secure a publicly exposed S3 bucket containing sensitive data",
"difficulty_multiplier": 1.0,
"max_steps": 15,
"ground_truth_timeline": [
"Public S3 bucket detected",
"Credentials discovered in bucket",
"Public access blocked"
],
"required_flags": {"s3_public_identified", "credentials_found", "public_access_blocked"},
"phase_weights": {"investigation": 0.5, "containment": 0.3, "eradication": 0.1, "recovery": 0.1},
"mitre_techniques": ["T1530"], # Data from Cloud Storage Object
"hints": [
"Start by reviewing SOC alerts",
"Check S3 bucket policies for public access",
"Look for sensitive files in public buckets"
]
},
"medium": {
"name": "Credential Compromise Response",
"description": "Leaked credentials from S3 were used; trace and revoke all compromised access",
"difficulty_multiplier": 1.5,
"max_steps": 25,
"ground_truth_timeline": [
"Leaked credentials identified",
"Compromised IAM role discovered",
"Role detached from resources",
"Credentials rotated"
],
"required_flags": {"leaked_creds_identified", "compromised_role_found", "role_detached", "credentials_rotated"},
"phase_weights": {"investigation": 0.3, "containment": 0.4, "eradication": 0.2, "recovery": 0.1},
"mitre_techniques": ["T1078", "T1552"], # Valid Accounts, Credentials in Files
"hints": [
"Trace credential usage in CloudTrail",
"Identify which roles were compromised",
"Detach roles before rotating credentials"
]
},
"hard": {
"name": "Full Incident Response - Ransomware",
"description": "Compromised credentials led to IAM backdoor and ransomware deployment. Full IR required.",
"difficulty_multiplier": 2.0,
"max_steps": 40,
"ground_truth_timeline": [
"Ransomware indicators detected",
"Backdoor IAM role identified",
"Forensic snapshot taken",
"Compromised instance isolated",
"Backdoor role removed",
"All credentials rotated",
"Systems verified clean"
],
"required_flags": {
"ransomware_detected", "backdoor_identified", "forensic_snapshot_taken",
"instance_isolated", "backdoor_removed", "all_creds_rotated", "systems_verified"
},
"phase_weights": {"investigation": 0.25, "containment": 0.25, "eradication": 0.3, "recovery": 0.2},
"mitre_techniques": ["T1486", "T1098", "T1078"], # Data Encrypted, Account Manipulation
"hints": [
"Preserve evidence with snapshots before isolation",
"Look for recently created IAM roles with admin access",
"Check for encryption activity on instances"
]
}
}
# =============================================================================
# MAIN ENVIRONMENT CLASS
# =============================================================================
class CloudSOCEnv(gym.Env):
"""
OpenEnv-CloudSOC Gymnasium Environment
Implements all 12 required mechanics for LLM agent evaluation on
cloud security incident response tasks.
"""
metadata = {"render_modes": ["human", "json"]}
def __init__(
self,
task: str = "easy",
seed: Optional[int] = None,
render_mode: str = "json",
initial_state: Optional[CloudState] = None,
verbose: bool = False
):
"""
Initialize the CloudSOC environment.
Args:
task: Difficulty level ("easy", "medium", "hard")
seed: Random seed for deterministic behavior (Mechanic #9)
render_mode: Output format ("human" or "json")
initial_state: Pre-existing state for multi-task campaigns (Mechanic #11)
verbose: Enable detailed logging
"""
super().__init__()
self.task = task
self.seed_value = seed
self.render_mode = render_mode
self.verbose = verbose
self.scenario = SCENARIOS[task]
# Deterministic seeding (Mechanic #9)
if seed is not None:
random.seed(seed)
self._deterministic = True
else:
self._deterministic = False
# Initialize or inherit state (Mechanic #11)
if initial_state is not None:
self.state = initial_state
self._inherited_state = True
else:
self.state = CloudState()
self._inherited_state = False
# Environment tracking
self.current_step = 0
self.max_steps = self.scenario["max_steps"]
self.total_reward = 0.0
self.rewards_history: List[float] = []
self.action_history: List[Dict] = []
self.done = False
self.last_action_error: Optional[str] = None
# Phase scoring (Mechanic #8)
self.phase_scores = {
"investigation": 0.0,
"containment": 0.0,
"eradication": 0.0,
"recovery": 0.0
}
# Tool usage analytics
self.tool_usage: Dict[str, int] = {}
self.query_costs = 0.0
# Gymnasium spaces (simplified for LLM)
self.action_space = spaces.Text(max_length=2048)
self.observation_space = spaces.Text(max_length=8192)
# Generate initial scenario
self._generate_scenario()
def _generate_scenario(self):
"""Generate the cloud environment state based on task difficulty"""
if self._inherited_state and self.task != "easy":
# Inherit and extend state for campaign continuity
self._extend_scenario_state()
else:
# Fresh state generation
self._generate_fresh_state()
# Generate logs with noise (Mechanic #1)
self._generate_logs()
# Generate alerts
self._generate_alerts()
def _generate_fresh_state(self):
"""Generate brand new cloud state"""
# Generate instance IDs deterministically if seeded
instance_suffix = f"{random.randint(0, 0xffffffff):08x}"
sg_suffix = f"{random.randint(0, 0xffffffff):08x}"
# Create EC2 instances
compromised_instance = EC2Instance(
instance_id=f"i-{instance_suffix}",
state=InstanceState.RUNNING,
attached_roles=["WebServerRole"],
security_groups=[f"sg-{sg_suffix}"],
is_compromised=self.task in ["medium", "hard"],
metadata={"name": "web-server-prod", "environment": "production"}
)
self.state.instances[compromised_instance.instance_id] = compromised_instance
# Add benign instances (noise)
for i in range(3):
suffix = f"{random.randint(0, 0xffffffff):08x}"
inst = EC2Instance(
instance_id=f"i-{suffix}",
state=InstanceState.RUNNING,
attached_roles=[f"ServiceRole{i}"],
security_groups=[f"sg-{random.randint(0, 0xffffffff):08x}"],
is_compromised=False,
metadata={"name": f"service-{i}", "environment": "production"}
)
self.state.instances[inst.instance_id] = inst
# Create IAM roles
web_role = IAMRole(
role_name="WebServerRole",
policies=["AmazonS3ReadOnly", "CloudWatchLogsFullAccess"],
is_compromised=self.task in ["medium", "hard"]
)
self.state.roles["WebServerRole"] = web_role
if self.task == "hard":
backdoor_role = IAMRole(
role_name="LegacyAdminBackup",
policies=["AdministratorAccess"],
is_compromised=True,
has_backdoor=True
)
self.state.roles["LegacyAdminBackup"] = backdoor_role
# Create S3 buckets
leaky_bucket = S3Bucket(
bucket_name="company-backup-2024",
is_public=True,
contains_credentials=True,
leaked_keys=["AKIAIOSFODNN7EXAMPLE"]
)
self.state.buckets["company-backup-2024"] = leaky_bucket
# Benign buckets
for name in ["company-logs", "company-assets", "company-config"]:
self.state.buckets[name] = S3Bucket(bucket_name=name, is_public=False)
# Security groups
self.state.security_groups[f"sg-{sg_suffix}"] = {
"id": f"sg-{sg_suffix}",
"name": "web-server-sg",
"inbound_rules": [
{"port": 443, "source": "0.0.0.0/0"},
{"port": 22, "source": "10.0.0.0/8"}
],
"is_isolated": False
}
# Leaked credentials
self.state.credentials["AKIAIOSFODNN7EXAMPLE"] = {
"key_id": "AKIAIOSFODNN7EXAMPLE",
"user": "backup-service",
"is_active": True,
"is_compromised": True,
"last_used": self._generate_timestamp(-2)
}
# Set ground truth timeline
self.state.ground_truth_timeline = self.scenario["ground_truth_timeline"]
def _extend_scenario_state(self):
"""Extend inherited state for harder scenarios"""
if self.task == "medium":
# Mark credentials as actively being abused
for cred in self.state.credentials.values():
if cred.get("is_compromised"):
cred["abuse_detected"] = True
cred["abuse_actions"] = ["AssumeRole", "ListBuckets", "GetObject"]
elif self.task == "hard":
# Add ransomware indicators
for inst in self.state.instances.values():
if inst.is_compromised:
inst.metadata["ransomware_indicators"] = True
inst.metadata["encrypted_files_detected"] = True
# Add backdoor role if not exists
if "LegacyAdminBackup" not in self.state.roles:
self.state.roles["LegacyAdminBackup"] = IAMRole(
role_name="LegacyAdminBackup",
policies=["AdministratorAccess"],
is_compromised=True,
has_backdoor=True
)
# Update ground truth
self.state.ground_truth_timeline = self.scenario["ground_truth_timeline"]
def _generate_logs(self):
"""Generate logs with deceptive elements (Mechanic #1)"""
base_time = datetime.utcnow() - timedelta(hours=24)
# Attack indicators (true positives)
attack_logs = [
LogEntry(
timestamp=self._generate_timestamp(-23),
source="CloudTrail",
message="AssumeRole called by AKIAIOSFODNN7EXAMPLE from IP 185.220.101.42",
severity="WARN",
is_attack_indicator=True
),
LogEntry(
timestamp=self._generate_timestamp(-22),
source="S3",
message="GetObject on s3://company-backup-2024/credentials.env from IP 185.220.101.42",
severity="WARN",
is_attack_indicator=True
),
LogEntry(
timestamp=self._generate_timestamp(-20),
source="GuardDuty",
message="UnauthorizedAccess:IAMUser/MaliciousIPCaller.Custom",
severity="CRITICAL",
is_attack_indicator=True
),
]
if self.task == "hard":
attack_logs.extend([
LogEntry(
timestamp=self._generate_timestamp(-18),
source="CloudTrail",
message="CreateRole 'LegacyAdminBackup' with AdministratorAccess by compromised session",
severity="CRITICAL",
is_attack_indicator=True
),
LogEntry(
timestamp=self._generate_timestamp(-15),
source="EC2",
message="Suspicious process execution: /tmp/.hidden/encrypt.sh on i-* instances",
severity="CRITICAL",
is_attack_indicator=True
),
])
# Red herrings (Mechanic #1 - Deceptive Environment)
red_herrings = [
LogEntry(
timestamp=self._generate_timestamp(-21),
source="CloudTrail",
message="Failed AssumeRole from IP 10.0.5.20 (security scanner - scheduled)",
severity="WARN",
is_red_herring=True
),
LogEntry(
timestamp=self._generate_timestamp(-19),
source="IAM",
message="Credential rotation reminder for service account 'deploy-bot'",
severity="INFO",
is_red_herring=True
),
LogEntry(
timestamp=self._generate_timestamp(-17),
source="EC2",
message="High CPU usage on i-abcd1234 (known batch job - analytics)",
severity="WARN",
is_red_herring=True
),
]
# Noise (benign activity)
noise_logs = []
for i in range(15):
noise_logs.append(LogEntry(
timestamp=self._generate_timestamp(-random.randint(1, 23)),
source=random.choice(["CloudTrail", "VPC", "Lambda", "ELB"]),
message=random.choice([
f"Successful authentication from IP 10.0.{random.randint(0,255)}.{random.randint(0,255)}",
f"Lambda function 'data-processor-{random.randint(1,5)}' completed in {random.randint(100,5000)}ms",
f"Health check passed for target group tg-{random.randint(1000,9999)}",
f"VPC flow log: ACCEPT from 10.0.0.{random.randint(1,255)} to 10.0.1.{random.randint(1,255)}",
f"Auto-scaling event: DesiredCapacity changed from {random.randint(2,5)} to {random.randint(2,5)}",
]),
severity="INFO",
is_noise=True
))
# Combine and shuffle
all_logs = attack_logs + red_herrings + noise_logs
random.shuffle(all_logs)
self.state.logs = sorted(all_logs, key=lambda x: x.timestamp, reverse=True)
def _generate_alerts(self):
"""Generate SOC alerts"""
self.state.alerts = [
Alert(
alert_id="ALT-001",
title="S3 Bucket Public Access Detected",
severity="HIGH",
source="AWS Config",
timestamp=self._generate_timestamp(-24),
details={"bucket": "company-backup-2024", "rule": "s3-bucket-public-read-prohibited"},
is_true_positive=True
),
Alert(
alert_id="ALT-002",
title="Unusual API Activity",
severity="MEDIUM",
source="GuardDuty",
timestamp=self._generate_timestamp(-22),
details={"finding_type": "UnauthorizedAccess:IAMUser/MaliciousIPCaller"},
is_true_positive=True
),
# False positive alert
Alert(
alert_id="ALT-003",
title="Brute Force Attempt Detected",
severity="LOW",
source="GuardDuty",
timestamp=self._generate_timestamp(-20),
details={"source_ip": "10.0.5.20", "note": "Internal security scanner"},
is_true_positive=False
),
]
if self.task == "hard":
self.state.alerts.append(Alert(
alert_id="ALT-004",
title="Ransomware Indicators Detected",
severity="CRITICAL",
source="CloudWatch",
timestamp=self._generate_timestamp(-15),
details={"indicators": ["mass file encryption", "ransom note creation"]},
is_true_positive=True
))
def _generate_timestamp(self, hours_offset: int) -> str:
"""Generate ISO timestamp with offset from now"""
dt = datetime.utcnow() + timedelta(hours=hours_offset)
return dt.strftime("%Y-%m-%dT%H:%M:%SZ")
def reset(self, seed: Optional[int] = None, options: Optional[Dict] = None) -> Tuple[str, Dict]:
"""Reset environment to initial state"""
if seed is not None:
self.seed_value = seed
random.seed(seed)
self.current_step = 0
self.total_reward = 0.0
self.rewards_history = []
self.action_history = []
self.done = False
self.last_action_error = None
self.phase_scores = {k: 0.0 for k in self.phase_scores}
self.tool_usage = {}
self.query_costs = 0.0
# Regenerate state
if not self._inherited_state:
self.state = CloudState()
self._generate_scenario()
else:
# Reset investigation state but keep infrastructure
self.state.discovered_flags = set()
self.state.completed_actions = []
self.state.agent_timeline = []
self.state.incident_phase = IncidentPhase.INVESTIGATION
observation = self._get_observation()
info = self._get_info()
return observation, info
def step(self, action: str) -> Tuple[str, float, bool, bool, Dict]:
"""
Execute one step in the environment.
Args:
action: JSON string containing tool call
Returns:
observation, reward, terminated, truncated, info
"""
self.current_step += 1
self.last_action_error = None
reward = 0.0
terminated = False
truncated = False
# Parse and validate action (Mechanic #7)
try:
action_dict = json.loads(action)
tool_call = ToolCall(**action_dict)
except json.JSONDecodeError as e:
self.last_action_error = f"PARSE_ERROR: Invalid JSON - {str(e)}"
reward = -0.02
return self._finish_step(reward, terminated, truncated)
except Exception as e:
self.last_action_error = f"VALIDATION_ERROR: {str(e)}"
reward = -0.02
return self._finish_step(reward, terminated, truncated)
# Track tool usage
self.tool_usage[tool_call.tool] = self.tool_usage.get(tool_call.tool, 0) + 1
# Record action
self.action_history.append({
"step": self.current_step,
"tool": tool_call.tool,
"args": tool_call.args,
"thought": tool_call.thought
})
# Execute tool and get result
result, step_reward, is_terminal, error = self._execute_tool(tool_call)
reward = step_reward
if error:
self.last_action_error = error
if is_terminal:
terminated = True
# Check for max steps (truncation)
if self.current_step >= self.max_steps:
truncated = True
self.done = terminated or truncated
return self._finish_step(reward, terminated, truncated)
def _finish_step(self, reward: float, terminated: bool, truncated: bool) -> Tuple[str, float, bool, bool, Dict]:
"""Finalize step and return results"""
self.total_reward += reward
self.rewards_history.append(reward)
observation = self._get_observation()
info = self._get_info()
return observation, reward, terminated, truncated, info
def _execute_tool(self, tool_call: ToolCall) -> Tuple[Dict, float, bool, Optional[str]]:
"""
Execute a tool call and return result.
Returns:
(result_dict, reward, is_terminal, error_message)
"""
tool = tool_call.tool
args = tool_call.args
# Tool dispatch
tool_handlers = {
'aws.cloudwatch.query_basic': self._handle_cloudwatch_basic,
'aws.cloudwatch.query_deep': self._handle_cloudwatch_deep,
'aws.ec2.isolate': self._handle_ec2_isolate,
'aws.ec2.terminate': self._handle_ec2_terminate,
'aws.ec2.describe': self._handle_ec2_describe,
'aws.ec2.snapshot': self._handle_ec2_snapshot,
'aws.iam.detach_role': self._handle_iam_detach,
'aws.iam.revoke_credentials': self._handle_iam_revoke,
'aws.iam.describe_role': self._handle_iam_describe,
'aws.iam.list_policies': self._handle_iam_list_policies,
'aws.s3.get_bucket_policy': self._handle_s3_get_policy,
'aws.s3.block_public_access': self._handle_s3_block_public,
'aws.s3.list_objects': self._handle_s3_list_objects,
'aws.rds.rotate_credentials': self._handle_rds_rotate,
'aws.security_group.modify': self._handle_sg_modify,
'aws.investigate': self._handle_investigate,
'aws.soc.close_incident': self._handle_close_incident,
'aws.soc.get_alerts': self._handle_get_alerts,
'aws.guardduty.get_findings': self._handle_guardduty,
'aws.cloudtrail.lookup_events': self._handle_cloudtrail,
'aws.config.get_compliance': self._handle_config_compliance,
'aws.ssm.run_command': self._handle_ssm_command,
'aws.lambda.list_functions': self._handle_lambda_list,
'aws.sts.get_caller_identity': self._handle_sts_identity,
}
handler = tool_handlers.get(tool)
if not handler:
return {}, -0.01, False, f"UNKNOWN_TOOL: {tool}"
return handler(args)
# =========================================================================
# TOOL HANDLERS
# =========================================================================
def _handle_cloudwatch_basic(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Basic CloudWatch query - truncated results, low cost (Mechanic #2)"""
cost = -0.01 # Resource cost
self.query_costs += abs(cost)
# Return only first 5 logs, summarized
relevant_logs = [
{"timestamp": log.timestamp, "source": log.source, "severity": log.severity}
for log in self.state.logs[:5]
]
# Discover basic flags
new_flags = set()
for log in self.state.logs[:5]:
if log.is_attack_indicator and "S3" in log.source:
new_flags.add("s3_activity_detected")
flag_reward = self._process_discovered_flags(new_flags)
return {
"status": "success",
"logs": relevant_logs,
"total_count": len(self.state.logs),
"note": "Truncated results. Use query_deep for full details."
}, cost + flag_reward, False, None
def _handle_cloudwatch_deep(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Deep CloudWatch query - full results, higher cost (Mechanic #2)"""
cost = -0.05 # Higher resource cost
self.query_costs += abs(cost)
# Return full log details
relevant_logs = [
{
"timestamp": log.timestamp,
"source": log.source,
"message": log.message,
"severity": log.severity
}
for log in self.state.logs
if not log.is_noise # Filter pure noise but keep red herrings
]
# Can discover more flags with deep query
new_flags = set()
for log in self.state.logs:
if log.is_attack_indicator:
if "S3" in log.source or "GetObject" in log.message:
new_flags.add("s3_public_identified")
if "credentials" in log.message.lower():
new_flags.add("credentials_found")
if "AssumeRole" in log.message and "185.220" in log.message:
new_flags.add("leaked_creds_identified")
if "CreateRole" in log.message:
new_flags.add("backdoor_identified")
if "ransomware" in log.message.lower() or "encrypt" in log.message.lower():
new_flags.add("ransomware_detected")
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["investigation"] += 0.1 if new_flags else 0.0
return {
"status": "success",
"logs": relevant_logs,
"total_count": len(relevant_logs)
}, cost + flag_reward, False, None
def _handle_ec2_describe(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Describe EC2 instances"""
instance_id = args.get("instance_id")
if instance_id:
inst = self.state.instances.get(instance_id)
if not inst:
return {"status": "error"}, -0.01, False, f"INSTANCE_NOT_FOUND: {instance_id}"
instances = {instance_id: self._serialize_instance(inst)}
else:
instances = {
iid: self._serialize_instance(inst)
for iid, inst in self.state.instances.items()
}
# Discover compromised instance flag
new_flags = set()
for iid, inst in self.state.instances.items():
if inst.is_compromised and inst.metadata.get("ransomware_indicators"):
new_flags.add("ransomware_detected")
flag_reward = self._process_discovered_flags(new_flags)
return {"status": "success", "instances": instances}, flag_reward, False, None
def _serialize_instance(self, inst: EC2Instance) -> Dict:
"""Serialize instance to dict"""
return {
"instance_id": inst.instance_id,
"state": inst.state.value,
"attached_roles": inst.attached_roles,
"security_groups": inst.security_groups,
"has_forensic_snapshot": inst.has_forensic_snapshot,
"metadata": inst.metadata
}
def _handle_ec2_snapshot(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Create forensic snapshot of instance"""
instance_id = args.get("instance_id")
if not instance_id:
return {}, -0.01, False, "MISSING_PARAM: instance_id required"
inst = self.state.instances.get(instance_id)
if not inst:
return {}, -0.01, False, f"INSTANCE_NOT_FOUND: {instance_id}"
if inst.state == InstanceState.TERMINATED:
return {}, -0.1, False, "ERROR: Cannot snapshot terminated instance"
inst.has_forensic_snapshot = True
self.state.completed_actions.append(f"snapshot:{instance_id}")
new_flags = {"forensic_snapshot_taken"} if inst.is_compromised else set()
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["containment"] += 0.15
self.state.agent_timeline.append(f"Forensic snapshot taken for {instance_id}")
return {
"status": "success",
"snapshot_id": f"snap-{hashlib.md5(instance_id.encode()).hexdigest()[:8]}",
"instance_id": instance_id
}, 0.05 + flag_reward, False, None
def _handle_ec2_isolate(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Isolate EC2 instance - requires snapshot first (Mechanic #3)"""
instance_id = args.get("instance_id")
if not instance_id:
return {}, -0.01, False, "MISSING_PARAM: instance_id required"
inst = self.state.instances.get(instance_id)
if not inst:
return {}, -0.01, False, f"INSTANCE_NOT_FOUND: {instance_id}"
# PRECONDITION CHECK (Mechanic #3)
if not inst.has_forensic_snapshot:
return {}, -0.05, False, "PRECONDITION_FAILED: Must take forensic snapshot before isolation"
if inst.state == InstanceState.ISOLATED:
return {"status": "already_isolated"}, 0.0, False, None
inst.state = InstanceState.ISOLATED
self.state.completed_actions.append(f"isolate:{instance_id}")
new_flags = {"instance_isolated"} if inst.is_compromised else set()
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["containment"] += 0.2
self.state.agent_timeline.append(f"Instance {instance_id} isolated")
return {
"status": "success",
"instance_id": instance_id,
"new_state": "isolated"
}, 0.1 + flag_reward, False, None
def _handle_ec2_terminate(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""ADVERSARIAL TRAP: Terminate instance destroys evidence (Mechanic #4)"""
instance_id = args.get("instance_id")
if not instance_id:
return {}, -0.01, False, "MISSING_PARAM: instance_id required"
inst = self.state.instances.get(instance_id)
if not inst:
return {}, -0.01, False, f"INSTANCE_NOT_FOUND: {instance_id}"
if inst.is_compromised:
# CATASTROPHIC ERROR - Evidence destruction
inst.state = InstanceState.TERMINATED
return {
"status": "terminated",
"warning": "CRITICAL: All forensic data has been permanently destroyed"
}, -1.0, True, "CRITICAL: FORENSIC_DATA_LOSS_TERMINATION"
# Non-compromised instance termination is just expensive
inst.state = InstanceState.TERMINATED
return {"status": "terminated"}, -0.1, False, None
def _handle_iam_detach(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Detach IAM role from resources"""
role_name = args.get("role_name")
if not role_name:
return {}, -0.01, False, "MISSING_PARAM: role_name required"
role = self.state.roles.get(role_name)
if not role:
return {}, -0.01, False, f"ROLE_NOT_FOUND: {role_name}"
if role.is_detached:
return {"status": "already_detached"}, 0.0, False, None
role.is_detached = True
self.state.completed_actions.append(f"detach_role:{role_name}")
new_flags = set()
if role.is_compromised:
new_flags.add("role_detached")
if role.has_backdoor:
new_flags.add("backdoor_removed")
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["eradication"] += 0.15
self.state.agent_timeline.append(f"Role {role_name} detached")
return {
"status": "success",
"role_name": role_name,
"detached": True
}, 0.05 + flag_reward, False, None
def _handle_iam_revoke(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Revoke IAM credentials"""
key_id = args.get("key_id") or args.get("resource_id")
if not key_id:
return {}, -0.01, False, "MISSING_PARAM: key_id required"
cred = self.state.credentials.get(key_id)
if not cred:
return {}, -0.01, False, f"CREDENTIAL_NOT_FOUND: {key_id}"
cred["is_active"] = False
cred["revoked_at"] = self._generate_timestamp(0)
new_flags = {"credentials_rotated"} if cred.get("is_compromised") else set()
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["eradication"] += 0.1
self.state.agent_timeline.append(f"Credentials {key_id} revoked")
return {
"status": "success",
"key_id": key_id,
"revoked": True
}, 0.05 + flag_reward, False, None
def _handle_iam_describe(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Describe IAM role"""
role_name = args.get("role_name")
if not role_name:
# List all roles
roles = {name: self._serialize_role(role) for name, role in self.state.roles.items()}
return {"status": "success", "roles": roles}, 0.0, False, None
role = self.state.roles.get(role_name)
if not role:
return {}, -0.01, False, f"ROLE_NOT_FOUND: {role_name}"
new_flags = set()
if role.is_compromised:
new_flags.add("compromised_role_found")
if role.has_backdoor:
new_flags.add("backdoor_identified")
flag_reward = self._process_discovered_flags(new_flags)
return {
"status": "success",
"role": self._serialize_role(role)
}, flag_reward, False, None
def _serialize_role(self, role: IAMRole) -> Dict:
return {
"role_name": role.role_name,
"policies": role.policies,
"is_detached": role.is_detached,
"credentials_rotated": role.credentials_rotated
}
def _handle_iam_list_policies(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""List IAM policies"""
all_policies = set()
for role in self.state.roles.values():
all_policies.update(role.policies)
return {"status": "success", "policies": list(all_policies)}, 0.0, False, None
def _handle_s3_get_policy(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Get S3 bucket policy"""
bucket_name = args.get("bucket_name")
if not bucket_name:
return {}, -0.01, False, "MISSING_PARAM: bucket_name required"
bucket = self.state.buckets.get(bucket_name)
if not bucket:
return {}, -0.01, False, f"BUCKET_NOT_FOUND: {bucket_name}"
new_flags = set()
if bucket.is_public:
new_flags.add("s3_public_identified")
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["investigation"] += 0.05 if new_flags else 0.0
return {
"status": "success",
"bucket_name": bucket_name,
"is_public": bucket.is_public,
"public_access_blocked": bucket.public_access_blocked,
"policy": {"Effect": "Allow", "Principal": "*"} if bucket.is_public else {}
}, flag_reward, False, None
def _handle_s3_block_public(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Block public access to S3 bucket"""
bucket_name = args.get("bucket_name")
if not bucket_name:
return {}, -0.01, False, "MISSING_PARAM: bucket_name required"
bucket = self.state.buckets.get(bucket_name)
if not bucket:
return {}, -0.01, False, f"BUCKET_NOT_FOUND: {bucket_name}"
if bucket.public_access_blocked:
return {"status": "already_blocked"}, 0.0, False, None
bucket.public_access_blocked = True
bucket.is_public = False
self.state.completed_actions.append(f"block_public:{bucket_name}")
new_flags = {"public_access_blocked"} if bucket.contains_credentials else set()
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["containment"] += 0.15
self.state.agent_timeline.append(f"Public access blocked for {bucket_name}")
return {
"status": "success",
"bucket_name": bucket_name,
"public_access_blocked": True
}, 0.1 + flag_reward, False, None
def _handle_s3_list_objects(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""List S3 bucket objects"""
bucket_name = args.get("bucket_name")
if not bucket_name:
return {}, -0.01, False, "MISSING_PARAM: bucket_name required"
bucket = self.state.buckets.get(bucket_name)
if not bucket:
return {}, -0.01, False, f"BUCKET_NOT_FOUND: {bucket_name}"
objects = ["backup-2024-01.tar.gz", "config.yaml", "logs/"]
if bucket.contains_credentials:
objects.append("credentials.env")
new_flags = set()
if bucket.contains_credentials:
new_flags.add("credentials_found")
flag_reward = self._process_discovered_flags(new_flags)
return {
"status": "success",
"bucket_name": bucket_name,
"objects": objects
}, flag_reward, False, None
def _handle_rds_rotate(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Rotate RDS credentials - requires role detachment first (Mechanic #3)"""
resource_id = args.get("resource_id")
if not resource_id:
return {}, -0.01, False, "MISSING_PARAM: resource_id required"
# PRECONDITION: Check if compromised roles are detached
compromised_attached = any(
role.is_compromised and not role.is_detached
for role in self.state.roles.values()
)
if compromised_attached:
return {}, -0.05, False, "PRECONDITION_FAILED: Detach compromised IAM roles before rotating credentials"
self.state.completed_actions.append(f"rotate_creds:{resource_id}")
# Check if all credentials are now rotated
all_rotated = all(
not cred.get("is_active", True)
for cred in self.state.credentials.values()
if cred.get("is_compromised")
)
new_flags = {"all_creds_rotated"} if all_rotated else set()
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["recovery"] += 0.15
self.state.agent_timeline.append(f"Credentials rotated for {resource_id}")
return {
"status": "success",
"resource_id": resource_id,
"credentials_rotated": True
}, 0.05 + flag_reward, False, None
def _handle_sg_modify(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Modify security group"""
sg_id = args.get("security_group_id")
action = args.get("action")
if not sg_id or not action:
return {}, -0.01, False, "MISSING_PARAM: security_group_id and action required"
sg = self.state.security_groups.get(sg_id)
if not sg:
return {}, -0.01, False, f"SECURITY_GROUP_NOT_FOUND: {sg_id}"
if action == "isolate":
sg["is_isolated"] = True
sg["inbound_rules"] = []
self.phase_scores["containment"] += 0.1
elif action == "restore":
sg["is_isolated"] = False
self.phase_scores["recovery"] += 0.05
return {"status": "success", "security_group_id": sg_id, "action": action}, 0.02, False, None
def _handle_investigate(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""General investigation tool"""
target_type = args.get("target_type")
target_id = args.get("target_id")
if not target_type or not target_id:
return {}, -0.01, False, "MISSING_PARAM: target_type and target_id required"
result = {"status": "success", "target_type": target_type, "target_id": target_id}
new_flags = set()
if target_type == "ip" and "185.220" in target_id:
result["findings"] = {
"reputation": "malicious",
"associated_campaigns": ["Credential Theft", "Cloud Exploitation"],
"geo": "TOR Exit Node"
}
new_flags.add("malicious_ip_confirmed")
elif target_type == "user":
result["findings"] = {"last_activity": self._generate_timestamp(-1)}
elif target_type == "role":
role = self.state.roles.get(target_id)
if role:
result["findings"] = self._serialize_role(role)
if role.is_compromised:
new_flags.add("compromised_role_found")
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["investigation"] += 0.05 if new_flags else 0.0
return result, flag_reward, False, None
def _handle_get_alerts(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Get SOC alerts"""
alerts = [
{
"alert_id": a.alert_id,
"title": a.title,
"severity": a.severity,
"source": a.source,
"timestamp": a.timestamp,
"details": a.details
}
for a in self.state.alerts
]
return {"status": "success", "alerts": alerts, "count": len(alerts)}, 0.0, False, None
def _handle_guardduty(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Get GuardDuty findings"""
findings = [
{
"id": f"gd-{i}",
"type": log.message.split(":")[0] if ":" in log.message else "UnauthorizedAccess",
"severity": log.severity,
"description": log.message
}
for i, log in enumerate(self.state.logs)
if log.source == "GuardDuty" and log.is_attack_indicator
]
new_flags = set()
if findings:
new_flags.add("guardduty_reviewed")
flag_reward = self._process_discovered_flags(new_flags)
return {"status": "success", "findings": findings}, flag_reward, False, None
def _handle_cloudtrail(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Lookup CloudTrail events"""
events = [
{
"timestamp": log.timestamp,
"event_name": log.message.split()[0] if log.message else "Unknown",
"source_ip": random.choice(MALICIOUS_IPS) if log.is_attack_indicator else random.choice(BENIGN_IPS),
"user": "compromised-user" if log.is_attack_indicator else "system",
"details": log.message
}
for log in self.state.logs
if log.source == "CloudTrail"
]
return {"status": "success", "events": events}, 0.0, False, None
def _handle_config_compliance(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Get AWS Config compliance status"""
rules = [
{"rule": "s3-bucket-public-read-prohibited", "compliance": "NON_COMPLIANT" if any(b.is_public for b in self.state.buckets.values()) else "COMPLIANT"},
{"rule": "iam-root-access-key-check", "compliance": "COMPLIANT"},
{"rule": "ec2-instance-no-public-ip", "compliance": "COMPLIANT"},
{"rule": "encrypted-volumes", "compliance": "COMPLIANT"},
]
new_flags = set()
if any(r["compliance"] == "NON_COMPLIANT" for r in rules):
new_flags.add("compliance_issue_found")
flag_reward = self._process_discovered_flags(new_flags)
return {"status": "success", "compliance_rules": rules}, flag_reward, False, None
def _handle_ssm_command(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Run SSM command on instance (for verification)"""
instance_id = args.get("instance_id")
command = args.get("command", "")
if not instance_id:
return {}, -0.01, False, "MISSING_PARAM: instance_id required"
inst = self.state.instances.get(instance_id)
if not inst:
return {}, -0.01, False, f"INSTANCE_NOT_FOUND: {instance_id}"
if inst.state != InstanceState.RUNNING:
return {}, -0.01, False, f"INSTANCE_NOT_RUNNING: {instance_id} is {inst.state.value}"
# Simulate command output
if "check" in command.lower() or "verify" in command.lower():
if inst.is_compromised and inst.metadata.get("ransomware_indicators"):
output = "WARNING: Suspicious processes detected\n/tmp/.hidden/encrypt.sh (running)"
else:
output = "System check passed. No anomalies detected."
new_flags = {"systems_verified"} if not inst.is_compromised else set()
flag_reward = self._process_discovered_flags(new_flags)
self.phase_scores["recovery"] += 0.1
return {"status": "success", "output": output}, flag_reward, False, None
return {"status": "success", "output": "Command executed"}, 0.0, False, None
def _handle_lambda_list(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""List Lambda functions"""
functions = [
{"name": "data-processor", "runtime": "python3.9", "last_modified": self._generate_timestamp(-48)},
{"name": "api-handler", "runtime": "nodejs18.x", "last_modified": self._generate_timestamp(-72)},
{"name": "auth-validator", "runtime": "python3.9", "last_modified": self._generate_timestamp(-24)},
]
return {"status": "success", "functions": functions}, 0.0, False, None
def _handle_sts_identity(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""Get current caller identity"""
return {
"status": "success",
"account": "123456789012",
"arn": "arn:aws:iam::123456789012:user/security-analyst",
"user_id": "AIDAEXAMPLEID"
}, 0.0, False, None
def _handle_close_incident(self, args: Dict) -> Tuple[Dict, float, bool, Optional[str]]:
"""
Close incident with timeline (Mechanic #12)
This is the FINAL action to complete a task successfully.
"""
timeline = args.get("timeline", [])
incident_id = args.get("incident_id", "INC-001")
if not timeline:
return {}, -0.1, False, "MISSING_PARAM: timeline array required"
if not isinstance(timeline, list) or len(timeline) < 2:
return {}, -0.1, False, "INVALID_PARAM: timeline must be an array with at least 2 events"
# Store agent's timeline
self.state.agent_timeline = timeline
# Grade the timeline (Mechanic #12)
timeline_score = self._grade_timeline(timeline)
# Check if required flags are discovered
required_flags = self.scenario["required_flags"]
discovered = self.state.discovered_flags
flags_complete = required_flags.issubset(discovered)
# Calculate final scores
final_scores = self.calculate_final_score()
# Determine success
success = flags_complete and timeline_score >= 0.5
self.state.incident_phase = IncidentPhase.CLOSED
return {
"status": "incident_closed",
"incident_id": incident_id,
"timeline_submitted": timeline,
"timeline_score": round(timeline_score, 2),
"flags_complete": flags_complete,
"discovered_flags": list(discovered),
"required_flags": list(required_flags),
"final_scores": final_scores,
"success": success
}, timeline_score * 0.5, True, None
def _grade_timeline(self, agent_timeline: List[str]) -> float:
"""
Grade the agent's incident timeline against ground truth.
Uses multi-factor scoring:
1. Keyword matching for event detection
2. Order preservation bonus
3. Completeness bonus
4. Penalty for extraneous/wrong events
"""
ground_truth = self.state.ground_truth_timeline
if not ground_truth:
return 0.5 # Default score if no ground truth
if not agent_timeline:
return 0.0
score = 0.0
max_score = len(ground_truth)
matched_indices = []
agent_lower = [t.lower() for t in agent_timeline]
# Score each ground truth event
for truth_idx, truth_event in enumerate(ground_truth):
truth_lower = truth_event.lower()
truth_words = set(truth_lower.split())
best_match_score = 0.0
best_match_idx = -1
for agent_idx, agent_event in enumerate(agent_lower):
agent_words = set(agent_event.split())
# Jaccard similarity
intersection = len(truth_words & agent_words)
union = len(truth_words | agent_words)
similarity = intersection / union if union > 0 else 0
# Keyword bonus for critical terms
critical_keywords = {"snapshot", "isolate", "detach", "rotate", "blocked", "revoke", "backdoor", "ransomware", "credential"}
keyword_matches = sum(1 for kw in critical_keywords if kw in agent_event)
keyword_bonus = min(0.2, keyword_matches * 0.1)
match_score = similarity + keyword_bonus
if match_score > best_match_score:
best_match_score = match_score
best_match_idx = agent_idx
if best_match_score >= 0.3: # Threshold for acceptable match
score += best_match_score
matched_indices.append((truth_idx, best_match_idx))
# Order preservation bonus (check if matched events maintain relative order)
if len(matched_indices) >= 2:
order_preserved = all(
matched_indices[i][1] < matched_indices[i+1][1]
for i in range(len(matched_indices) - 1)
if matched_indices[i][1] != -1 and matched_indices[i+1][1] != -1
)
if order_preserved:
score += 0.5 # Bonus for correct ordering
# Completeness bonus
coverage = len([m for m in matched_indices if m[1] != -1]) / max_score
score += coverage * 0.3
# Normalize to 0-1 range
normalized = min(1.0, score / (max_score + 0.8))
# Scale to 0.1 - 0.9 per requirements
return 0.1 + (normalized * 0.8)
def _process_discovered_flags(self, new_flags: set) -> float:
"""
Process newly discovered flags and return reward.
Implements Mechanic #5 (Gradient Scoring).
"""
actually_new = new_flags - self.state.discovered_flags
self.state.discovered_flags.update(actually_new)
# Reward per new flag (Mechanic #5)
reward = 0.02 * len(actually_new)
return reward
def calculate_final_score(self) -> Dict[str, float]:
"""
Calculate final score breakdown (Mechanic #8).
Returns structured scoring across IR phases.
"""
weights = self.scenario["phase_weights"]
# Normalize phase scores to 0-1 range
max_scores = {
"investigation": 0.5,
"containment": 0.6,
"eradication": 0.4,
"recovery": 0.3
}
normalized = {}
for phase, score in self.phase_scores.items():
max_s = max_scores.get(phase, 0.5)
normalized[phase] = min(1.0, score / max_s) if max_s > 0 else 0.0
# Weighted average
weighted_total = sum(
normalized[phase] * weights[phase]
for phase in weights
)
# Adjust weighted_total to range 0.1 to 0.9 rigidly
weighted_total = 0.1 + (weighted_total * 0.8)
return {
"investigation": round(0.1 + (normalized["investigation"] * 0.8), 3),
"containment": round(0.1 + (normalized["containment"] * 0.8), 3),
"eradication": round(0.1 + (normalized["eradication"] * 0.8), 3),
"recovery": round(0.1 + (normalized["recovery"] * 0.8), 3),
"weighted_total": round(weighted_total, 3),
"query_costs": round(self.query_costs, 3),
"total_reward": round(self.total_reward, 3)
}
def _get_observation(self) -> str:
"""Get current observation as JSON string"""
obs = {
"step": self.current_step,
"phase": self.state.incident_phase.value,
"discovered_flags": list(self.state.discovered_flags),
"last_error": self.last_action_error,
"alerts_pending": len([a for a in self.state.alerts if a.is_true_positive]),
"instances_count": len(self.state.instances),
"compromised_instances": len([i for i in self.state.instances.values() if i.is_compromised]),
}
if self.action_history:
last_action = self.action_history[-1]
obs["last_action"] = {
"tool": last_action["tool"],
"result": "error" if self.last_action_error else "success"
}
return json.dumps(obs, indent=2)
def _get_info(self) -> Dict:
"""Get additional info dict"""
return {
"task": self.task,
"scenario": self.scenario["name"],
"current_step": self.current_step,
"max_steps": self.max_steps,
"total_reward": round(self.total_reward, 4),
"discovered_flags": list(self.state.discovered_flags),
"required_flags": list(self.scenario["required_flags"]),
"phase_scores": self.phase_scores.copy(),
"tool_usage": self.tool_usage.copy(),
"last_action_error": self.last_action_error
}
def render(self, mode: Optional[str] = None) -> Optional[str]:
"""Render current state"""
mode = mode or self.render_mode
if mode == "json":
return json.dumps({
"task": self.task,
"step": self.current_step,
"state": {
"instances": {k: self._serialize_instance(v) for k, v in self.state.instances.items()},
"discovered_flags": list(self.state.discovered_flags),
"phase": self.state.incident_phase.value
},
"scores": self.calculate_final_score()
}, indent=2)
elif mode == "human":
lines = [
f"=== CloudSOC Environment ===",
f"Task: {self.scenario['name']} ({self.task})",
f"Step: {self.current_step}/{self.max_steps}",
f"Phase: {self.state.incident_phase.value}",
f"Flags: {len(self.state.discovered_flags)}/{len(self.scenario['required_flags'])}",
f"Total Reward: {self.total_reward:.4f}",
f"Last Error: {self.last_action_error or 'None'}"
]
return "\n".join(lines)
return None
def close(self):
"""Clean up environment"""
pass
def get_state_for_next_task(self) -> CloudState:
"""
Export current state for multi-task campaign (Mechanic #11).
Returns the state object to pass to the next task's environment.
"""
return self.state
def get_system_prompt(self) -> str:
"""
Generate the system prompt for the LLM agent.
Implements Mechanic #10 (Chain-of-Thought Prompting).
"""
available_tools = [
"aws.cloudwatch.query_basic - Query logs (truncated, cost: -0.01)",
"aws.cloudwatch.query_deep - Query logs (full, cost: -0.05)",
"aws.ec2.describe - Describe EC2 instances",
"aws.ec2.snapshot - Create forensic snapshot (REQUIRED before isolation)",
"aws.ec2.isolate - Isolate compromised instance",
"aws.ec2.terminate - ⚠️ DANGER: Permanently terminates instance, destroys evidence",
"aws.iam.describe_role - Describe IAM role",
"aws.iam.detach_role - Detach IAM role from resources",
"aws.iam.revoke_credentials - Revoke compromised credentials",
"aws.iam.list_policies - List IAM policies",
"aws.s3.get_bucket_policy - Check S3 bucket policy",
"aws.s3.block_public_access - Block public access to bucket",
"aws.s3.list_objects - List bucket contents",
"aws.rds.rotate_credentials - Rotate DB credentials (requires role detachment first)",
"aws.security_group.modify - Modify security group (isolate/restore)",
"aws.investigate - Investigate IP, user, role, or instance",
"aws.soc.get_alerts - Get current SOC alerts",
"aws.guardduty.get_findings - Get GuardDuty findings",
"aws.cloudtrail.lookup_events - Lookup CloudTrail events",
"aws.config.get_compliance - Get AWS Config compliance status",
"aws.ssm.run_command - Run command on instance for verification",
"aws.lambda.list_functions - List Lambda functions",
"aws.sts.get_caller_identity - Get current caller identity",
"aws.soc.close_incident - Close incident with timeline (FINAL action)"
]
# Get progress hints based on current state
progress_hints = []
discovered = self.state.discovered_flags
required = self.scenario["required_flags"]
remaining = required - discovered
if len(discovered) == 0:
progress_hints.append("Start by gathering information: check alerts and logs")
elif len(remaining) <= 2:
progress_hints.append("You're close! Complete remaining actions and close the incident")
return f"""You are an expert Cloud Security Analyst responding to a security incident in AWS.
## INCIDENT: {self.scenario['name']}
{self.scenario['description']}
## RESPONSE FORMAT (MANDATORY)
You MUST respond with valid JSON in this exact format:
```json
{{
"thought": "Your 1-sentence reasoning about what to do next and why",
"tool": "tool_name",
"args": {{"param": "value"}}
}}
```
## ⚠️ CRITICAL WARNINGS - READ CAREFULLY
1. **NEVER use aws.ec2.terminate on compromised instances** - This permanently destroys all forensic evidence and fails the mission
2. **ALWAYS take a forensic snapshot BEFORE isolating** - Isolation without snapshot = precondition failure
3. **ALWAYS detach compromised IAM roles BEFORE rotating credentials** - Rotation without detachment = precondition failure
4. **Query costs matter**: basic=-0.01, deep=-0.05 - Balance thoroughness with efficiency
## INCIDENT RESPONSE PROCEDURE
1. **Investigation**: Gather information (alerts, logs, findings)
2. **Containment**: Stop the spread (snapshot → isolate, block access)
3. **Eradication**: Remove threat (detach roles, revoke credentials)
4. **Recovery**: Restore operations (rotate credentials, verify systems)
5. **Close**: Submit timeline with aws.soc.close_incident
## AVAILABLE TOOLS
{chr(10).join(f'• {t}' for t in available_tools)}
## TOOL PARAMETER SCHEMAS (MANDATORY - Use exact parameter names)
• aws.s3.get_bucket_policy: args: {{"bucket_name": "string"}}
• aws.s3.list_objects: args: {{"bucket_name": "string"}}
• aws.s3.block_public_access: args: {{"bucket_name": "string"}}
• aws.ec2.describe: args: {{"instance_id": "string"}} (optional)
• aws.ec2.isolate: args: {{"instance_id": "string"}}
• aws.ec2.snapshot: args: {{"instance_id": "string"}}
• aws.ec2.terminate: args: {{"instance_id": "string"}}
• aws.iam.describe_role: args: {{"role_name": "string"}}
• aws.iam.detach_role: args: {{"role_name": "string", "resource_id": "string"}}
• aws.iam.revoke_credentials: args: {{"principal": "string", "credential_type": "string"}}
• aws.cloudwatch.query_basic: args: {{"log_group": "string", "query": "string"}}
• aws.cloudwatch.query_deep: args: {{"log_group": "string", "query": "string"}}
• aws.cloudtrail.lookup_events: args: {{"event_names": [list], "max_results": int}}
• aws.soc.close_incident: args: {{"timeline": [list of strings]}}
• aws.investigate: args: {{"query": "string"}}
• aws.security_group.modify: args: {{"group_id": "string", "action": "string"}}
• aws.rds.rotate_credentials: args: {{"instance_id": "string"}}
• aws.guardduty.get_findings: args: {{}} (no args required)
• aws.soc.get_alerts: args: {{}} (no args required)
• aws.config.get_compliance: args: {{}} (no args required)
• aws.ssm.run_command: args: {{"instance_id": "string", "command": "string"}}
• aws.lambda.list_functions: args: {{}} (no args required)
• aws.sts.get_caller_identity: args: {{}} (no args required)
## CURRENT STATUS
• Step: {self.current_step}/{self.max_steps}
• Phase: {self.state.incident_phase.value}
• Progress: {len(self.state.discovered_flags)}/{len(self.scenario['required_flags'])} flags discovered
• Query costs so far: {self.query_costs:.2f}
{chr(10).join(f'• Hint: {h}' for h in progress_hints)}
## FINAL STEP
To complete this incident, call `aws.soc.close_incident` with a timeline array summarizing the incident chronologically. Example:
```json
{{"thought": "Investigation complete, closing incident", "tool": "aws.soc.close_incident", "args": {{"timeline": ["S3 bucket found public", "Credentials discovered", "Access blocked"]}}}}
```
Think step-by-step. Each action should move you closer to resolution."""
# =============================================================================
# HELPER FUNCTIONS
# =============================================================================
def create_campaign() -> List[Tuple[str, Optional[CloudState]]]:
"""
Create a multi-task campaign (Mechanic #11).
Returns list of (task_name, initial_state) tuples.
"""
return [
("easy", None),
("medium", None), # Will receive state from easy
("hard", None), # Will receive state from medium
]
def run_campaign_task(
task: str,
initial_state: Optional[CloudState] = None,
seed: Optional[int] = None
) -> Tuple[CloudSOCEnv, CloudState]:
"""
Run a single task in the campaign.
Returns (env, final_state) for chaining to next task.
"""
env = CloudSOCEnv(task=task, initial_state=initial_state, seed=seed)
return env, env.state
# =============================================================================
# MAIN (for testing)
# =============================================================================
if __name__ == "__main__":
# Test environment creation
env = CloudSOCEnv(task="easy", seed=42)
obs, info = env.reset()
print("=== CloudSOC Environment Test ===")
print(f"Task: {env.task}")
print(f"Scenario: {env.scenario['name']}")
print(f"Max Steps: {env.max_steps}")
print(f"\nInitial Observation:\n{obs}")
print(f"\nSystem Prompt:\n{env.get_system_prompt()[:500]}...")
# Test a simple action
test_action = json.dumps({
"thought": "First, I need to check the SOC alerts to understand the incident",
"tool": "aws.soc.get_alerts",
"args": {}
})
obs, reward, term, trunc, info = env.step(test_action)
print(f"\n=== After Step 1 ===")
print(f"Reward: {reward}")
print(f"Observation:\n{obs}")
env.close()
|