dmitrl-dev/aisecurity/agent-security

-[ ] Нет eval() или Function() в коде
-[ ] Нет exec() без whitelist
-[ ] subprocesses имеют timeout
-[ ] Sandbox для выполнения кода

-[ ] User input не конкатенируется с system prompt
-[ ] Есть input sanitization
-[ ] Metadata (filenames, URLs) санитизированы
-[ ] Vision inputs проверяются

-[ ] Approval system для опасных операций
-[ ] allowlist/denylist для команд
-[ ] Confirmation для file write/delete
-[ ] Network restrictions

-[ ] Rate limit на API endpoints
-[ ] Rate limit на exec calls
-[ ] Rate limit на LLM calls
-[ ] Throttling на resource-intensive operations

-[ ] Memory не содержит credentials
-[ ] Session data encrypted
-[ ] TTL на sensitive facts
-[ ] No cross-session data leakage

-[ ] Extensions cryptographically signed
-[ ] Skills verified before load
-[ ] No arbitrary code from untrusted sources
-[ ] Package integrity checks

-[ ] Principle of least privilege
-[ ] No sudo/root by default
-[ ] Limited file system scope
-[ ] Restricted network access

-[ ] All exec commands logged
-[ ] All file operations logged
-[ ] All network requests logged
-[ ] Immutable audit trail

-[ ] evaluateEnabled: false by default
-[ ] exec.ask: always by default
-[ ] No credentials in default paths
-[ ] Secure default configurations

-[ ] Human-in-the-loop для critical operations
-[ ] Clear approval workflows
-[ ] Ability to interrupt and rollback
-[ ] Transparent decision logging

1. Run skill on target codebase
2. Check each A01-A10 category
3. Generate findings JSON
4. Store in RLM (domain: agent-security)
5. Block if CRITICAL findings

{
  "target": "clawdbot",
  "standard": "OWASP Agentic Top 10 2026",
  "risk_level": "CRITICAL | HIGH | MEDIUM | LOW",
  "findings": [
    {
      "id": "AGENT-001",
      "category": "A01",
      "severity": "CRITICAL",
      "title": "eval() enabled by default",
      "file": "pw-tools-core.interactions.ts",
      "line": 227,
      "description": "Browser tool uses eval() with evaluateEnabled: true",
      "remediation": "Set evaluateEnabled: false as default",
      "cwe": "CWE-94"
    }
  ],
  "score": {
    "A01": 2,
    "A02": 5,
    "A03": 7,
    ...
  },
  "passed": 4,
  "failed": 6
}

# After audit completion
rlm_add_hierarchical_fact(
    content=f"Agent Security Audit: {target} - {risk_level}, {len(findings)} findings",
    level=2,
    domain="agent-security",
    ttl_days=365
)