File size: 3,260 Bytes
1ab3545 | 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 | {
"name": "Adversarial red-team loop",
"objective": "Discover and confirm agent-system failures inside an authorized threat model with reproducible evidence and safe disclosure.",
"trigger": {
"type": "scheduled",
"cadence_or_event": "A bounded pre-release campaign runs after a model, tool, permission, or policy change."
},
"intake": {
"sources": ["target version", "written threat model", "allowed attack classes", "regression corpus", "known findings", "disclosure contacts"],
"selection_rule": "Probe one uncovered, in-scope attack class at a time using synthetic data and a sandboxed target."
},
"workspace": {
"isolation": "Sandboxed target with synthetic accounts, fake credentials, network boundaries, and rate limits.",
"allowed_actions": ["generate in-scope probes", "run against the sandbox", "reproduce and minimize traces", "write private finding reports"],
"disallowed_actions": ["test production", "use real user data", "exfiltrate data", "persist access", "publish unpatched findings"]
},
"context": {
"required_files": ["threat model", "safety policy", "tool permission map", "disclosure procedure"],
"runtime_sources": ["target configuration", "probe and seed", "full agent trace", "tool calls", "regression results"]
},
"agents": [
{
"role": "Attacker",
"responsibility": "Generate and adapt probes inside the written scope and budget."
},
{
"role": "Reproducer",
"responsibility": "Confirm candidate failures independently in a clean sandbox."
},
{
"role": "Minimizer",
"responsibility": "Reduce confirmed behavior to the smallest safe, reliable test case."
},
{
"role": "Judge",
"responsibility": "Classify confirmed, duplicate, expected, or inconclusive behavior against the written policy."
}
],
"verification": {
"gates": ["behavior reproduces on the recorded target version", "finding cites a violated policy or boundary", "a separate judge confirms the result", "minimized test contains no real secrets or user data", "duplicate and severity review are complete"],
"receipts": ["target version and environment", "seed and probe", "full and minimized traces", "reproduction count", "policy citation", "severity and disclosure status"]
},
"state": {
"artifacts": ["private campaign ledger", "confirmed finding report", "regression corpus entry"],
"update_rule": "Record every candidate with target version, trace, reproduction status, classification, duplicate link, severity rationale, and disclosure owner."
},
"budget": {
"max_retries": 3,
"max_runtime_minutes": 120
},
"escalation": {
"conditions": ["critical behavior is confirmed", "production impact is possible", "scope is ambiguous", "containment is unsafe", "real data may be exposed"],
"destination": "Security owner through the private disclosure channel"
},
"exit": {
"success": "A campaign completes with every confirmed failure reproduced, minimized, privately reported, and added to the regression corpus.",
"stop_without_success": "Budget is exhausted, the target is unstable, scope becomes ambiguous, a rate limit is reached, or human intervention is required."
}
}
|