Security audit
Security checks for vulnerabilities and agentic risk
This is a disclosed Markdown-only guide for coordinating multiple AI agents; it can lead agents to modify files and run tests, but those behaviors fit its stated orchestration purpose.
Install only if you want your agent to coordinate multiple workers that may read and modify project files, run tests, and create `.orchestrator/` artifacts. Keep secrets blocked, approve any Bash/test/deploy commands before they run, and use explicit file scopes and budgets for each orchestration.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
max_budget_usd: 0.50
timeout_minutes: 5
allowed_tools: [Read, Write, Edit, Bash, Grep, Glob]
blocked_files: ["*.env", "*.key", "*.pem", "credentials.*"]
total:
max_agents: 8
max_budget_usd: 5.00
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
initial_prompt: |
Build a complete JWT auth system:
- Login endpoint with email/password
- JWT access tokens (15min) + refresh tokens (7 days)
- Rate limiting: 5 failed attempts → 15min lockout
- Secure token storage (httpOnly cookies)
- Token rotation on refresh
The Test stage tells the agent to write tests and run them, but provides no warning that command execution may consume resources, trigger project scripts, access networked dependencies, or alter the local environment. In this orchestration context, the omission is more dangerous because the pipeline is designed for automation, so users may adopt it verbatim and allow agents to execute commands with minimal review.
The conditional stage examples include deploy and rollback flows but omit any warning that these actions can affect production systems, data integrity, availability, or release state. Even as examples, presenting automated deploy/rollback logic in a production-grade orchestration skill without approval gates or environment restrictions can encourage unsafe copy-paste adoption.
The template explicitly instructs agents to write intermediate outputs into workspace files under .orchestrator/ without any warning, consent step, or scoping guidance. In a multi-agent orchestration skill, normalizing unattended file writes increases the chance of unintended workspace modification, clobbering existing files, or persisting sensitive data to disk.
No suspicious patterns detected.