This project currently supports the v0.7.x line.

Security fixes will be backported to the latest published v0.7 release where practical.

Security contact: Charles Zeller / charleszeller@gmail.com

If you discover a security issue, report it privately before public disclosure. Please include the affected version, a short summary, reproduction steps, and any relevant logs or indicators. Do not open a public issue with exploit details. If a GitHub Security Advisory process is available for the repository, use that. Otherwise, email the security contact above.

We aim to acknowledge good-faith reports promptly and will coordinate disclosure before any public write-up.

Safeloop provides governance primitives for local AI agent loops:

• policy gating before execution
• circuit breaking during execution
• action ledgers for review and auditability
• markdown reports for human review

It is not a sandbox and is not a complete security boundary.

This package does not replace:

• tool sandboxing
• credential isolation
• network egress controls
• file-system permissions
• prompt injection defenses
• manual diff review
• least-privilege access

• Run local-first whenever possible.
• Use least privilege for tools and credentials.
• Do not run agents with production credentials.
• Sandbox external tools and networked actions.
• Avoid exposing .env, SSH keys, npm tokens, GitHub tokens, or API keys in prompts, logs, or ledger entries.
• Require human approval for high-risk actions.
• Review git diff before commit or push.
• Treat the live simulation harness as proof of control logic, not as a security boundary.