Skill Auditor v2

Security scanner for OpenClaw skills. Detects malicious code, obfuscated payloads, prompt injection, social engineering, typosquatting, and data exfiltration...

This skill appears to be what it says: a Python-based auditor plus a shell quarantine helper. Before running it, review the bundled audit_skill.py (it will scan files and may fetch remote slugs) and quarantine.sh yourself, and run them in an isolated environment (container or VM) on untrusted skills. Be aware the --slug mode will perform network fetches to retrieve remote skills — if you need to avoid network I/O, run the auditor only against local directories. Confirm the default production directory in quarantine.sh matches your environment before approving any automatic copy/installation, and inspect audit-report.json results before installing any audited skill. The pre-scan prompt-injection flag is explained by the auditor including a list of injection signatures; this is expected but always worth a quick manual check because such patterns could be abused if the skill were modified by a malicious actor.

Type: OpenClaw Skill Name: skill-auditor-v2 Version: 2.0.0 The OpenClaw AgentSkills skill 'skill-auditor-v2' is a security scanner designed to detect malicious code, prompt injection, and data exfiltration in other skills. All analyzed files (code and documentation) are consistent with this stated purpose. The `audit_skill.py` script implements various detection layers, including static analysis, deobfuscation, and threat intelligence, but these are for *detecting* threats in *other* skills, not for performing malicious actions itself. The `quarantine.sh` script orchestrates the audit and offers conditional installation of *other* skills after explicit user confirmation. The use of `subprocess.run` to execute `clawhub inspect` for fetching remote skills is a legitimate function for an auditor and is implemented in a manner that mitigates direct shell injection (passing arguments as a list). No evidence of intentional harmful behavior, data exfiltration, backdoors, or prompt injection against the agent by *this* skill was found.