Tetra Scar Safety

Agent safety that learns from incidents. Reflex arc blocks repeat threats without LLM calls.

This skill appears to be what it claims: a local safety checker with an append-only incident log. Before installing, consider: (1) audit runs will read files you point it at — do not run audits over root or system directories unless you intend that; (2) safety_scars.jsonl (default ./safety_scars.jsonl) is append-only and may contain incident text that includes secrets or sensitive details — store it with restrictive file permissions, move it to a secure location, or encrypt it if needed; (3) the skill has no network or credential requirements in the provided code, but you should still review the source included in the package before enabling autonomous invocation; (4) for CI/pre-commit use, configure scar file location and retention policy to avoid leaking historical incidents. If you want higher assurance, run the included tests locally and inspect audit/recorded entries to confirm no unexpected behavior.

Type: OpenClaw Skill Name: tetra-scar-safety Version: 0.1.0 The 'scar-safety' skill is a security utility designed to protect AI agents by detecting and blocking dangerous commands, secret exposures, and potential injection attacks. It implements a two-layer defense: a static regex-based scanner (targeting patterns like 'rm -rf', 'DROP TABLE', and leaked AWS/GitHub keys) and a 'scar' memory system that records past incidents to prevent recurrence via keyword matching. The code in 'scar_safety.py' is transparent, uses only Python standard library modules, and includes a comprehensive test suite ('test_scar_safety.py'). No evidence of data exfiltration, unauthorized remote execution, or malicious prompt injection was found; the tool's high-privilege actions (like directory auditing) are strictly aligned with its stated purpose as a safety framework.