OpenClaw Skill Auditor

Scans ClawHub skills to detect malicious code, obfuscated payloads, and social engineering via pattern matching, deobfuscation, and LLM analysis before insta...

This skill appears to implement a legitimate scanner, but review before use: (1) be aware it expects (but doesn't declare) the 'clawhub' CLI to fetch remote skills and optionally an LLM CLI (Gemini) for deeper analysis — install and verify those tools from trusted sources first; (2) run the auditor in an isolated sandbox or VM when scanning untrusted skills because it fetches remote skill packages; (3) the script writes collected suspicious snippets to a predictable /tmp file and suggests sending them to an LLM or emailing a personal address — do NOT send secrets or plaintext sensitive files to third-party LLMs or personal emails; (4) if you plan to use the LLM analysis path, configure it to use a trusted, privacy-preserving endpoint or do the analysis offline; and (5) ask the author to explicitly declare required binaries/dependencies and to remove or make the reporting workflow safer (use a private/internal reporting channel). If you need higher assurance, run the script on a throwaway environment and manually inspect its output before following any 'send' recommendations.

Type: OpenClaw Skill Name: openclaw-skill-auditor Version: 1.0.0 The OpenClaw AgentSkills skill bundle 'skill-auditor' is designed to detect malicious code and security threats in other ClawHub skills. Its `SKILL.md` and `scripts/audit.sh` implement pattern matching, deobfuscation, and LLM analysis to identify high-risk behaviors like base64/curl piped to bash, known C2 servers (e.g., 91.92.242.30), credential theft keywords, and access to sensitive directories. All identified high-risk capabilities are for detection purposes, not execution by the skill itself. The script uses temporary files and `clawhub inspect` for its legitimate auditing function, and the LLM analysis step explicitly saves suspicious code locally and provides a prompt suggestion for the *user* to analyze it, rather than instructing the agent to perform malicious actions or exfiltrate data.