Agent Security Audit

Security skill для аудита AI-агентов на базе исследования Clawdbot

Применимость

Используется для аудита любых AI-агентов с:

• Terminal/exec access
• Browser automation
• File system access
• Memory/context persistence
• External integrations (Telegram, WhatsApp, Slack)

OWASP Agentic Top 10 2026 Checklist

A01: Uncontrolled Tool Execution

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

Severity: 🔴 CRITICAL

A02: Prompt Injection

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

Severity: 🔴 CRITICAL

A03: Missing Tool Guardrails

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

Severity: 🔴 CRITICAL

A04: No Rate Limiting

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

Severity: 🟠 HIGH

A05: Insecure Memory/Context

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

Severity: 🟠 HIGH

A06: Missing Extension Signatures

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

Severity: 🟡 MEDIUM

A07: Excessive Permissions

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

Severity: 🟠 HIGH

A08: No Audit Logging

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

Severity: 🟡 MEDIUM

A09: Unsafe Defaults

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

Severity: 🔴 CRITICAL

A10: Insufficient Human Oversight

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

Severity: 🟠 HIGH

Workflow

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

Output Format

{
  "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
}

Integration with RLM

# 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
)

Related Skills

• security-audit — general code security
• two-stage-review — spec + quality review
• tdd-enforcement — test-first development