Skip to content

INCIDENT_RESPONSE.md

Latest commit

 

History

History
63 lines (44 loc) · 1.98 KB

INCIDENT_RESPONSE.md

File metadata and controls

63 lines (44 loc) · 1.98 KB

Incident Response Guide

Phase: 7 (Enterprise Reality) Status: Initial draft.

Detecting a Potential Bypass

Signs that the kernel may have been bypassed or is misconfigured:

  1. Agent performed an action without a corresponding PERMITTED audit entry → Agent is not routed through the verifier. Fix: wire all tool calls through verify().

  2. audit_log.verify_chain() returns False → Audit log was tampered with or entries were deleted. → Preserve the log, treat all decisions since the last verified entry as untrustworthy.

  3. Machine acted on a resource it has no claim for → Either the action was not verified, or the registry was modified post-freeze. → Audit registry mutation calls.

Immediate Response: Revoke a Compromised Agent

# Immediate revocation of a single agent
registry.revoke_all("CompromisedBot")

# Cascading revocation: revoke CompromisedBot and all agents it delegated to
count = registry.revoke_cascading("CompromisedBot")
print(f"Revoked {count} claims across delegation subtree")

Forensics: Audit Log Analysis

log = AuditLog(path="/var/log/kernel.jsonl")
# Load from file (re-read for forensics):
import json
entries = [json.loads(line) for line in open("/var/log/kernel.jsonl")]

# Find all PERMITTED actions by a specific actor
suspect_permits = [e for e in entries if e["permitted"] and "SuspectBot" in e["action_id"]]

# Verify chain integrity
if not log.verify_chain():
    print("ALERT: Audit log tampered — escalate to security team")

Escalation Criteria

Event Severity Action
verify_chain() fails Critical Preserve logs, revoke all affected agents, escalate
PERMITTED with no prior claim Critical Agent bypassed gate — investigate orchestration layer
Unexpected BLOCKED volume spike High May indicate attack probing; review violations
Machine claimed sovereignty flag High Review action source; check for prompt injection