Spaces:
Running
Running
File size: 15,161 Bytes
ca83593 | 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 | #!/usr/bin/env python3
"""
IAMSentinel Baseline Inference Script
======================================
Runs a GPT-4o ReAct agent against all 3 tasks and reports scores.
Usage:
export OPENAI_API_KEY=sk-...
python scripts/baseline_agent.py [--task all|task1|task2|task3] [--seed 42] [--model gpt-4o]
Reproducible baseline scores (seed=42, complexity=medium, model=gpt-4o-mini):
Task 1 (Easy): ~0.55β0.70
Task 2 (Medium): ~0.35β0.50
Task 3 (Hard): ~0.20β0.35
"""
import argparse
import json
import os
import sys
import time
from typing import Optional
try:
from openai import OpenAI
except ImportError:
print("ERROR: openai package not installed. Run: pip install openai")
sys.exit(1)
# Ensure package is importable
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from iamsentinel import IAMSentinelEnv
# ββββββββββββββββββββββββββββββββββββββββββββββ
# System prompt for the ReAct agent
# ββββββββββββββββββββββββββββββββββββββββββββββ
SYSTEM_PROMPT = """You are an expert cloud security analyst specialising in AWS IAM security.
You are operating inside a simulated IAM environment and must complete security tasks.
You interact with the environment by outputting JSON actions. Each response must contain
EXACTLY ONE action as a JSON block in this format:
```json
{
"action": "<action_name>",
... action parameters ...
}
```
Available actions:
1. list_principals β {"action": "list_principals", "kind": "all"|"user"|"role"}
2. list_policies β {"action": "list_policies", "principal_arn": "<arn or null>"}
3. get_policy β {"action": "get_policy", "policy_arn": "<arn>"}
4. get_principal β {"action": "get_principal", "principal_arn": "<arn>"}
5. get_role_trust β {"action": "get_role_trust", "role_arn": "<arn>"}
6. query_audit_log β {"action": "query_audit_log", "filter": {"event_name": "...", "severity": "...", "principal_arn": "...", "source_ip": "..."}, "limit": 20}
7. trace_escalation_path β {"action": "trace_escalation_path", "from_principal_arn": "<arn>", "to_principal_arn": null}
8. flag_finding β {
"action": "flag_finding",
"finding_type": "wildcard_policy"|"mfa_disabled"|"stale_admin_role"|"privilege_escalation_path"|"exposed_trust_policy"|"suspicious_event",
"affected_principal_arn": "<arn or null>",
"affected_policy_arn": "<arn or null>",
"severity": "low"|"medium"|"high"|"critical",
"description": "<description>",
"mitre_technique": "<T-code or null>",
"evidence": ["<arn or event_id>", ...]
}
9. remediate β {"action": "remediate", "remediation_type": "detach_policy"|"delete_user"|"require_mfa"|"update_trust_policy", "target_arn": "<arn>", "policy_arn": "<arn or null>"}
10. attribute_attack β {
"action": "attribute_attack",
"compromised_principal_arn": "<arn>",
"attack_technique": "<description>",
"mitre_techniques": ["T1078.004", ...],
"lateral_movement_path": ["<arn1>", "<arn2>"],
"containment_actions": ["disable_user:<arn>", "delete_function:<name>", ...]
}
Strategy guidelines:
- For Task 1: List all principals and their policies. Check for wildcards, MFA, stale roles, exposed trust policies.
- For Task 2: Find principals with iam:PassRole. Trace escalation paths. Look for lambda + createUser chains.
- For Task 3: Query audit logs by severity=critical first, then trace suspicious sequences. Look for CreateFunctionβCreateUser chains from unusual IPs.
Be systematic. Think step by step before each action. Flag findings as you discover them.
For Task 3, finish with attribute_attack once you've gathered enough evidence.
"""
# ββββββββββββββββββββββββββββββββββββββββββββββ
# JSON action parser
# ββββββββββββββββββββββββββββββββββββββββββββββ
def extract_json_action(text: str) -> Optional[dict]:
"""Extract the first JSON block from model output."""
import re
# Try fenced code block first
pattern = r"```(?:json)?\s*(\{.*?\})\s*```"
match = re.search(pattern, text, re.DOTALL)
if match:
try:
return json.loads(match.group(1))
except json.JSONDecodeError:
pass
# Try raw JSON
pattern2 = r"\{[^{}]*\"action\"[^{}]*\}"
match2 = re.search(pattern2, text, re.DOTALL)
if match2:
try:
return json.loads(match2.group(0))
except json.JSONDecodeError:
pass
# Try to find largest JSON object
for start in range(len(text)):
if text[start] == "{":
for end in range(len(text), start, -1):
if text[end-1] == "}":
try:
obj = json.loads(text[start:end])
if "action" in obj:
return obj
except json.JSONDecodeError:
continue
return None
def obs_to_text(obs_dict: dict, step: int) -> str:
"""Convert observation dict to a concise text summary for the LLM."""
parts = [f"[Step {step}] Budget remaining: {obs_dict.get('budget_remaining', '?')}"]
if obs_dict.get("hints"):
parts.append("Hints: " + " | ".join(obs_dict["hints"]))
if obs_dict.get("findings"):
parts.append(f"Findings so far ({len(obs_dict['findings'])}):")
for f in obs_dict["findings"][-3:]: # last 3
parts.append(f" - [{f['severity']}] {f['finding_type']}: {f['description'][:80]}")
if obs_dict.get("principals"):
parts.append(f"Principals returned: {len(obs_dict['principals'])}")
for p in obs_dict["principals"][:5]:
mfa = "βMFA" if p.get("mfa_enabled") else "βMFA"
parts.append(
f" {p['kind']}: {p['name']} | {mfa} | "
f"last_active={p['last_active_days']}d | "
f"policies={len(p.get('policies', []))}"
)
if len(obs_dict["principals"]) > 5:
parts.append(f" ... and {len(obs_dict['principals'])-5} more")
if obs_dict.get("policies"):
parts.append(f"Policies returned: {len(obs_dict['policies'])}")
for p in obs_dict["policies"][:5]:
wildcard = "β WILDCARD" if p.get("is_wildcard") else ""
parts.append(f" {p['name']} {wildcard} | arn={p['arn']}")
if p.get("statements"):
actions = p["statements"][0].get("actions", [])
parts.append(f" actions: {actions[:5]}")
if len(obs_dict["policies"]) > 5:
parts.append(f" ... and {len(obs_dict['policies'])-5} more")
if obs_dict.get("audit_events"):
parts.append(f"Audit events returned: {len(obs_dict['audit_events'])}")
for e in obs_dict["audit_events"][:8]:
parts.append(
f" [{e.get('severity','?')}] {e['event_time']} | "
f"{e['event_name']} | {e['principal_name']} | ip={e['source_ip']}"
)
if len(obs_dict["audit_events"]) > 8:
parts.append(f" ... and {len(obs_dict['audit_events'])-8} more")
if obs_dict.get("escalation_paths"):
parts.append(f"Escalation paths found: {len(obs_dict['escalation_paths'])}")
for ep in obs_dict["escalation_paths"][:3]:
parts.append(f" Path (risk={ep.get('risk_score','?')}): {' β '.join(ep['path'])}")
if obs_dict.get("role_trust_policy"):
parts.append(f"Trust policy: {json.dumps(obs_dict['role_trust_policy'], indent=2)[:300]}")
if obs_dict.get("done"):
parts.append("EPISODE DONE.")
return "\n".join(parts)
# ββββββββββββββββββββββββββββββββββββββββββββββ
# Agent runner
# ββββββββββββββββββββββββββββββββββββββββββββββ
def run_agent(
task_id: str,
seed: int = 42,
model: str = "gpt-4o-mini",
complexity: str = "medium",
verbose: bool = True,
) -> dict:
api_key = os.environ.get("OPENAI_API_KEY")
if not api_key:
raise ValueError("OPENAI_API_KEY environment variable not set")
client = OpenAI(api_key=api_key)
env = IAMSentinelEnv(task_id=task_id, seed=seed, complexity=complexity)
obs = env.reset()
task_cfg = {
"task1": {"name": "Misconfiguration Scanner", "difficulty": "Easy"},
"task2": {"name": "Privilege Escalation Path Detection","difficulty": "Medium"},
"task3": {"name": "Live Attack Attribution", "difficulty": "Hard"},
}[task_id]
if verbose:
print(f"\n{'='*60}")
print(f"Task: {task_cfg['name']} ({task_cfg['difficulty']})")
print(f"Seed: {seed} | Model: {model} | Complexity: {complexity}")
print(f"{'='*60}")
# Build conversation history
messages = [{"role": "system", "content": SYSTEM_PROMPT}]
# Initial user message with task description
initial_msg = (
f"Task: {obs.task_description}\n\n"
f"Account ID: {obs.account_id}\n"
f"Max steps: {obs.max_steps}\n"
)
if obs.hints:
initial_msg += "\nHints:\n" + "\n".join(f"- {h}" for h in obs.hints)
initial_msg += "\n\nBegin your investigation. Output one JSON action."
messages.append({"role": "user", "content": initial_msg})
episode_done = False
step = 0
final_score = 0.0
total_reward = 0.0
action_history = []
while not episode_done and step < env._max_steps():
step += 1
# ββ Call LLM ββββββββββββββββββββββββββ
try:
response = client.chat.completions.create(
model=model,
messages=messages,
temperature=0.2,
max_tokens=800,
)
assistant_text = response.choices[0].message.content
except Exception as e:
print(f" [Step {step}] LLM error: {e}")
time.sleep(2)
continue
messages.append({"role": "assistant", "content": assistant_text})
# ββ Parse action βββββββββββββββββββββββ
action_dict = extract_json_action(assistant_text)
if action_dict is None:
if verbose:
print(f" [Step {step}] Could not parse action from: {assistant_text[:100]}")
feedback = "ERROR: Could not parse a valid JSON action. Output ONLY a JSON block."
messages.append({"role": "user", "content": feedback})
continue
action_name = action_dict.get("action", "unknown")
action_history.append(action_name)
if verbose:
print(f" [Step {step}] Action: {action_name}", end="")
key_params = {k: v for k, v in action_dict.items()
if k != "action" and v is not None}
if key_params:
print(f" | params: {json.dumps(key_params)[:100]}", end="")
print()
# ββ Step environment βββββββββββββββββββ
try:
next_obs, reward, done, info = env.step(action_dict)
except Exception as e:
feedback = f"ERROR executing action: {e}. Try a different action."
messages.append({"role": "user", "content": feedback})
continue
total_reward += reward.total
episode_done = done
if done and info.get("final_score") is not None:
final_score = info["final_score"]
if verbose:
print(f" [Step {step}] Episode done. Final score: {final_score:.3f}")
# ββ Build feedback message βββββββββββββ
obs_dict = next_obs.model_dump()
feedback_text = obs_to_text(obs_dict, step)
if reward.step_reward != 0:
feedback_text += f"\n[Reward signal: {reward.step_reward:+.3f}]"
if obs_dict.get("findings"):
feedback_text += f"\n[Total findings logged: {len(obs_dict['findings'])}]"
if not done:
feedback_text += "\n\nContinue your investigation. Output one JSON action."
messages.append({"role": "user", "content": feedback_text})
# Small delay to respect rate limits
time.sleep(0.3)
return {
"task_id": task_id,
"task_name": task_cfg["name"],
"difficulty": task_cfg["difficulty"],
"seed": seed,
"model": model,
"final_score": final_score,
"total_reward": total_reward,
"steps_taken": step,
"action_history":action_history,
"state": env.state(),
}
# ββββββββββββββββββββββββββββββββββββββββββββββ
# Main entry point
# ββββββββββββββββββββββββββββββββββββββββββββββ
def main():
parser = argparse.ArgumentParser(description="IAMSentinel Baseline Agent")
parser.add_argument("--task", default="all", help="task1|task2|task3|all")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--model", default="gpt-4o-mini")
parser.add_argument("--complexity", default="medium", help="easy|medium|hard")
parser.add_argument("--output", default=None, help="Save results to JSON file")
parser.add_argument("--quiet", action="store_true")
args = parser.parse_args()
tasks = ["task1", "task2", "task3"] if args.task == "all" else [args.task]
results = []
for task_id in tasks:
result = run_agent(
task_id=task_id,
seed=args.seed,
model=args.model,
complexity=args.complexity,
verbose=not args.quiet,
)
results.append(result)
# ββ Print summary ββββββββββββββββββββββββββ
print("\n" + "="*60)
print("BASELINE SCORES SUMMARY")
print("="*60)
print(f"{'Task':<35} {'Score':>6} {'Steps':>5} {'Difficulty'}")
print("-"*60)
for r in results:
print(
f"{r['task_name']:<35} {r['final_score']:>6.3f} "
f"{r['steps_taken']:>5} {r['difficulty']}"
)
print("-"*60)
avg = sum(r["final_score"] for r in results) / len(results)
print(f"{'Average':<35} {avg:>6.3f}")
print("="*60)
if args.output:
with open(args.output, "w") as f:
json.dump(results, f, indent=2)
print(f"\nResults saved to {args.output}")
return results
if __name__ == "__main__":
main()
|