Spaces:
Sleeping
Sleeping
File size: 9,375 Bytes
d064478 | 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 | """OpenEnv-style wrapper for the ShadowOps cybersecurity environment."""
from __future__ import annotations
from dataclasses import dataclass, field
import time
from pathlib import Path
from typing import Any
from agent_memory import ActionMemoryRecord, SessionMemory
from shadowops_env import ACTIONS, UniversalShadowEnv, compute_ambiguity, extract_features
from training.reward_rubric import score_reward_rubric
from training.shadowops_training_common import build_q_aware_decision, normalize_action_output
VALID_ACTIONS = tuple(ACTIONS.values())
ACTION_TO_INDEX = {label: index for index, label in ACTIONS.items()}
@dataclass
class EpisodeStep:
step: int
action: str
reward: float
domain: str
outcome: str
risk_score: float
decision_context: dict[str, Any] = field(default_factory=dict)
class ShadowOpsOpenEnv:
"""Small Gym/OpenEnv-compatible wrapper around ``UniversalShadowEnv``.
The wrapper keeps the model-free demo deterministic while exposing a
judge-friendly environment contract: ``reset()``, ``step(action)``,
``state()``, and ``close()``. Actions affect production/shadow state,
quarantine holds, memory, accumulated risk, and future observations.
"""
metadata = {
"name": "shadowops",
"render_modes": [],
"actions": VALID_ACTIONS,
}
def __init__(
self,
*,
seed: int = 42,
malicious_rate: float = 0.5,
episode_max_length: int = 8,
memory_path: Path | str | None = None,
persist_memory: bool = False,
) -> None:
self.seed = seed
self.episode_max_length = episode_max_length
self._env = UniversalShadowEnv(
malicious_rate=malicious_rate,
episode_max_length=episode_max_length,
mode="openenv",
seed=seed,
)
self.memory = SessionMemory(
persistence_enabled=persist_memory,
storage_path=memory_path or Path(__file__).resolve().parent / "data" / "openenv_session_memory.json",
)
self.session_id = f"openenv-{seed}"
self.history: list[EpisodeStep] = []
self._last_observation: dict[str, Any] | None = None
self._last_info: dict[str, Any] = {}
def reset(self) -> dict[str, Any]:
"""Reset the episode and return an observation object."""
obs_text, obs_vec = self._env.reset()
self.history.clear()
self._last_info = {}
self._last_observation = self._format_observation(obs_text, obs_vec)
return self._last_observation
def step(self, action: str | int) -> tuple[dict[str, Any], float, bool, dict[str, Any]]:
"""Apply an action and return ``(observation, reward, done, info)``."""
action_label = self._normalize_action(action)
scenario = dict(self._env._current_scenario or {})
risk_vector = extract_features(
scenario.get("domain", "SOC"),
scenario.get("intent", "UNKNOWN"),
scenario.get("raw_payload", ""),
self._env.rng,
)
risk_score = float(sum(risk_vector[:4]) / max(len(risk_vector[:4]), 1))
memory_context = self.memory.summarize_memory_context(self.session_id)
decision_context = build_q_aware_decision(
scenario.get("domain", "SOC"),
scenario.get("intent", "UNKNOWN"),
scenario.get("raw_payload", ""),
risk_vector,
actor="openenv_agent",
session_id=self.session_id,
service=scenario.get("domain", "unknown"),
environment="production",
provided_evidence=[],
timestamp=len(self.history) + 1,
memory_context=memory_context,
)
obs_text, obs_vec, reward, done, info = self._env.step(ACTION_TO_INDEX[action_label])
info = dict(info)
self.memory.add_record(
ActionMemoryRecord(
actor="openenv_agent",
session_id=self.session_id,
service=str(info.get("domain", scenario.get("domain", "unknown"))),
domain=str(info.get("domain", scenario.get("domain", "unknown"))),
environment="production",
timestamp=time.time(),
decision=action_label,
risk_score=float(decision_context.get("cumulative_risk_score", risk_score)),
action_summary=str(scenario.get("raw_payload", "")),
indicators=list(decision_context.get("risk_indicators", [])),
)
)
updated_memory_context = self.memory.summarize_memory_context(self.session_id)
rubric = score_reward_rubric(
{
"correct_action": decision_context.get("decision"),
"severity": "CRITICAL" if decision_context.get("cumulative_risk_score", 0.0) >= 0.80 else "MEDIUM",
"risk_score": decision_context.get("cumulative_risk_score", 0.0),
"raw_payload": scenario.get("raw_payload", ""),
"required_evidence": decision_context.get("required_evidence", []),
"provided_evidence": [],
"is_malicious": decision_context.get("decision") in {"BLOCK", "FORK", "QUARANTINE"},
},
action_label,
decision_context,
memory_context=updated_memory_context,
)
info.update(
{
"available_actions": list(VALID_ACTIONS),
"decision_context": decision_context,
"memory_context": updated_memory_context,
"reward_rubric": rubric,
"risk_score": decision_context.get("risk_score", 0.0),
"cumulative_risk_score": decision_context.get("cumulative_risk_score", 0.0),
"missing_evidence": decision_context.get("missing_evidence", []),
"evidence_plan": decision_context.get("evidence_plan", []),
"safe_outcome": decision_context.get("safe_outcome", ""),
}
)
self.history.append(
EpisodeStep(
step=int(info.get("step", len(self.history) + 1)),
action=action_label,
reward=float(reward),
domain=str(info.get("domain", "unknown")),
outcome=str(info.get("outcome", "unknown")),
risk_score=float(decision_context.get("cumulative_risk_score", 0.0)),
decision_context=decision_context,
)
)
self._last_info = info
self._last_observation = self._format_observation(obs_text, obs_vec)
return self._last_observation, float(reward), bool(done), info
def state(self) -> dict[str, Any]:
"""Return the current incident-response state without mutating it."""
memory_context = self.memory.summarize_memory_context(self.session_id)
return {
"session_id": self.session_id,
"step_count": self._env.step_count,
"episode_reward": self._env.episode_reward,
"available_actions": list(VALID_ACTIONS),
"production": self._env.get_production_snapshot(),
"health": self._env.get_health_scores(),
"forensic_log": self._env.get_forensic_log(),
"incident_reports": self._env.get_incident_reports(),
"memory_context": memory_context,
"history": [step.__dict__ for step in self.history],
"last_info": self._last_info,
}
def close(self) -> None:
"""Close hook for OpenEnv/Gym compatibility."""
return None
def clear_memory(self) -> None:
self.memory.clear()
def _format_observation(self, obs_text: str, obs_vec: list[float]) -> dict[str, Any]:
current = dict(self._env._current_scenario or {})
q_active = bool(obs_vec[16]) if len(obs_vec) > 16 else False
q_steps = obs_vec[17] if len(obs_vec) > 17 else 0.0
return {
"prompt": obs_text,
"risk_vector": list(obs_vec[:16]),
"quarantine": {
"active": q_active,
"steps_remaining_normalized": q_steps,
},
"available_actions": list(VALID_ACTIONS),
"incident_state": {
"domain": current.get("domain", "unknown"),
"intent": current.get("intent", "unknown"),
"payload": current.get("raw_payload", ""),
"tier": current.get("tier", "unknown"),
"ambiguity_score": compute_ambiguity(obs_vec[:16]),
"step_count": self._env.step_count,
"health": self._env.get_health_scores(),
"memory_context": self.memory.summarize_memory_context(self.session_id),
},
}
@staticmethod
def _normalize_action(action: str | int) -> str:
if isinstance(action, int):
if action not in ACTIONS:
raise ValueError(f"Invalid ShadowOps action index: {action}")
return ACTIONS[action]
parsed = normalize_action_output(str(action))
if parsed not in VALID_ACTIONS:
raise ValueError(f"Invalid ShadowOps action label: {action}")
return parsed
def make_env(**kwargs: Any) -> ShadowOpsOpenEnv:
return ShadowOpsOpenEnv(**kwargs)
|