AGENT SKILL

skillsentry-audit

Audit AI Agent Skills for malware before installation with SkillSentry, a fully-local open-source scanner that runs a local model. Use when the user asks to audit a skill, scan a skill, check if a skill is safe, detect malicious skills, review a SKILL.md package, or verify a skill from a registry such as ClawHub before installing it. Runs an 8-stage inspection: SHA-256 incremental cache, spec validation with typosquat detection, AST anchors fused with local-model semantic analysis, vision scanning for image-hidden prompt injection, sandboxed execution with context-aware honeypots, Poisoned Well active defense (fake credentials, exfil tracked to C2), sandbox-evasion detection, weighted verdict fusion. Returns a SAFE, SUSPICIOUS, or MALICIOUS verdict with a 0-100 risk score, line-accurate citations, and SARIF/STIX reports. Trigger keywords: audit skill, scan skill, skill security, malicious skill, pre-install check, SKILL.md review, supply-chain audit, typosquat check.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
1

01 前言 Front Matter

name
skillsentry-audit
description
Audit AI Agent Skills for malware before installation with SkillSentry, a fully-local open-source scanner that runs a local model. Use when the user asks to audit a skill, scan a skill, check if a skill is safe, detect malicious skills, review a SKILL.md package, or verify a skill from a registry such as ClawHub before installing it. Runs an 8-stage inspection: SHA-256 incremental cache, spec validation with typosquat detection, AST anchors fused with local-model semantic analysis, vision scanning for image-hidden prompt injection, sandboxed execution with context-aware honeypots, Poisoned Well active defense (fake credentials, exfil tracked to C2), sandbox-evasion detection, weighted verdict fusion. Returns a SAFE, SUSPICIOUS, or MALICIOUS verdict with a 0-100 risk score, line-accurate citations, and SARIF/STIX reports. Trigger keywords: audit skill, scan skill, skill security, malicious skill, pre-install check, SKILL.md review, supply-chain audit, typosquat check.

02 资源文件 1

03 SKILL.md 正文

SkillSentry Audit

Pre-install security gate for Agent Skills. Everything runs locally (Ollama + Docker + SQLite) — audited skill content never leaves the machine unless the user explicitly enabled the Gemini API fallback.

When to use

  • Before installing any skill from a registry (ClawHub, marketplaces) or a third-party repo.
  • When re-auditing installed skills after upstream updates (supply-chain poisoning check).
  • When the user asks: "is this skill safe?", "scan this skill", "audit SKILL.md", "check for typosquats".

How to run

Use the bundled script (preferred):

scripts/audit.sh <path-to-skill-directory> [extra skill-sentry-cli flags]

Equivalent direct CLI:

skill-sentry-cli scan <path-to-skill-directory> --fail-on suspicious

scripts/audit.sh exits 1 on any SUSPICIOUS or MALICIOUS verdict. It also exits 1 on a degraded scan, because a degraded scan is never SAFE: its verdict is at least SUSPICIOUS. Pass --fail-on malicious only after the user has reviewed a SUSPICIOUS result and decided to accept it.

Useful flags: --demo-mode (offline canned demo), --no-cache (force full re-scan), --sarif-out report.sarif, --format json, --backend local|cloud.

Interpreting results

Verdict Risk score Recommended action
✅ SAFE < 30 Install normally
⚠️ SUSPICIOUS 30–70 Review the cited file:line findings with the user before deciding
🚫 MALICIOUS > 70, or hard-evidence override (confidence 1.0) Do not install. Report the SARIF/STIX exports; note any C2 indicators

Always surface to the user: the verdict, risk score, top findings with file:line citations, whether the result was ⚡ cached (< 2 s), and — if triggered — the Poisoned Well summary (targeted assets, exfil channel, C2 indicators). Never suppress a MALICIOUS verdict, and never present a degraded scan (e.g., sandbox-unavailable) as a clean pass.

Requirements

  • skill-sentry-cli CLI installed (pip install skill-sentry-cli)
  • Ollama running with the local model pulled (skill-sentry-cli doctor names it) — or --backend cloud with GEMINI_API_KEY, or --demo-mode
  • Docker daemon available for sandbox stages (Stage 4/4.5/5); without it the scan degrades to static-only analysis and says so explicitly

Safety notes

  • SkillSentry never connects to attacker infrastructure; C2 indicators are harvested passively from the sinkhole.
  • All canary credentials served during analysis are fake by construction (marker SKS-<8hex>).
  • Audited packages are executed only inside the hardened sandbox (60 s, 512 MB, 1 CPU, sinkhole-only egress).

Local analysis by default · open weights. Cloud analysis is opt-in and is labelled as such.

04 同仓库更多 Skills