ClawTrust
2.6k Downloads
Overview
Audit locally installed agent skills for security and policy issues using the SkillLens CLI (skilllens scan / skilllens config) and produce a risk-focused report per skill.
Key Advantages
1.Provides structured, security-oriented audits of local skills (exfiltration, execution, persistence, privilege bypass, prompt injection, overbroad triggers).
2.Uses SkillLens to inventory skills and surface unsafe/suspicious verdicts for prioritized review.
3.Encourages conservative handling of gaps (missing auditor CLIs or skipped scans treated as “manual review required,” not “safe”).
4.Promotes safer workflows: avoiding default script execution, narrowing permissions, and adding confirmation gates.
5.Supports targeted scanning of specific directories or all configured roots, enabling both ad hoc and routine audits.
Use Cases
- Scanning a Codex or Claude skills directory (e.g., ~/.codex/skills) before enabling or updating skills in an agent environment.
- Auditing a repository of skills (e.g., skilllens scan ./skills) prior to sharing with a team or deploying to production.
- Performing periodic security reviews of installed skills to identify new high-risk capabilities or configuration drift.
- Investigating suspicious behavior by enumerating local skills and reviewing those with broad filesystem/network permissions or shell execution.
- Supporting compliance or policy reviews by generating per-skill risk scores, verdicts, and evidence-backed issue lists.
Evaluation Scores
8.3
/ 10
Reliability
7.5
Functionality
8.5
Usability
8.5
Safety
9.0
Performance
8.0
Compatibility
7.5
Based on 1 evaluation · Latest: 3/19/2026
Download Trend
Loading...
Evaluation History (1)
8.3/103/19/2026▼
OS: linux-x64LLM: moonshotai/kimi-k2.5
**Quick judgment**
SkillLens Audit is a well-scoped, security-focused skill for auditing locally installed agent skills. It relies on the external SkillLens CLI to scan directories, then guides a structured, conservative review process centered on realistic abuse scenarios.
**What it does well**
- Inventories skills via `skilllens scan [path] [--auditor claude|codex]`, including Codex/Claude-style skill directories.
- Focuses on concrete risks: exfiltration, code execution, persistence, privilege/approval bypass, prompt injection, and overbroad triggers.
- Enforces a cautious stance: missing auditor CLIs or skipped scans are treated as “manual review required,” not “safe.”
- Encourages reviewing SKILL.md plus referenced `scripts/`, `references/`, and `assets/` without executing scripts by default.
- Produces per-skill verdicts (safe/suspicious/unsafe), 0–100 risk scores, and evidence-based issue bullets with recommended mitigations.
**Key risks / limitations**
- **External dependency**: Requires the `skilllens` CLI (and optionally auditor CLIs like Claude/Codex); missing tools or misconfiguration can reduce coverage.
- **Coverage gaps**: Static/manual review may miss runtime behaviors, dynamic downloads, or indirect effects not evident from SKILL.md and bundled files.
- **False sense of safety**: Users might over-trust a “safe” verdict; the workflow assumes ongoing caution and manual review of higher-risk skills.
- **Environment-specific behavior**: Risk depends on how and where skills are executed (privileges, network access), which the auditor cannot fully infer from files alone.
**Recommended scenarios**
- Before enabling or updating skills in environments like Codex/Claude (`~/.codex/skills` or similar directories).
- Security and compliance reviews of skill repositories prior to team-wide deployment or open-source release.
- Periodic hygiene checks to detect skills with broad filesystem/network access, shell command execution, or external download behavior.
- Incident response investigations when an agent appears to behave unexpectedly and you need a prioritized list of risky local skills to inspect.
Comments (0)
No comments yet. Be the first!