Cisco AI Skill Scanner

该 Skill 封装 Cisco AI Skill Scanner，通过脚本或 CLI 扫描技能目录，使用静态分析、行为数据流和可选 LLM 分析检测提示注入、数据外泄、凭据窃取及恶意代码，输出严重级别与 Markdown/HTML 报告，可用于发布前安全审计。

godsboy

热度 36.05

最近更新 2026-05-11 17:38:19

分类：安全依赖：代码执行 / 读写文件 / 内部调用大模型应用场景：代码审查 / 报告撰写 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add cisco-skill-scanner

Metadata

name: skill-scanner
description: Scan agent skills for security threats using the Cisco AI skill-scanner CLI. Triggers on: scan skill for security, check skill safety, audit skill code, skill-scanner, detect prompt injection in skill, skill malware check, scan skills directory, security audit skill, validate skill before publishing, skill threat detection.
homepage: https://github.com/cisco-ai-defense/skill-scanner
license: MIT
metadata: {"openclaw":{"emoji":"🔍","requires":{"bins":["skill-scanner"]},"install":[{"id":"pip","kind":"run","label":"Install Cisco AI skill-scanner (pip)","run":"pip install cisco-ai-skill-scanner --break-system-packages"}]}}

Cisco AI Skill Scanner

Security scanner for agent skills. Detects prompt injection, data exfiltration, credential harvesting, and malicious code patterns using static analysis, behavioral dataflow, and optional LLM-as-a-judge.

Quick Scan (recommended default)

bash scripts/scan.sh /path/to/skill

Commands

Scan a single skill

skill-scanner scan /path/to/skill

Scan with behavioral analysis (dataflow)

skill-scanner scan /path/to/skill --use-behavioral

Full scan (all engines, requires API key)

SKILL_SCANNER_LLM_API_KEY="$ANTHROPIC_API_KEY" \
SKILL_SCANNER_LLM_MODEL="claude-sonnet-4-5" \
skill-scanner scan /path/to/skill --use-behavioral --use-llm --enable-meta --llm-provider anthropic

Scan all skills in a directory

skill-scanner scan-all /root/clawd/skills --recursive --use-behavioral

Scan with detailed markdown report

skill-scanner scan /path/to/skill --use-behavioral --format markdown --detailed

Scan before publishing to ClawHub

skill-scanner scan /path/to/skill --use-behavioral --fail-on-severity medium

Severity Levels

CRITICAL / HIGH — Do not install/publish. Review and fix immediately.
MEDIUM — Suspicious patterns. Investigate before use.
LOW — Minor issues. Document and accept risk or fix.
INFO — Informational only (e.g. missing license). Safe to publish.
SAFE (0 findings) — No known threat patterns detected.

Key Flags

Flag	Purpose
`--use-behavioral`	AST dataflow analysis (recommended, no API key needed)
`--use-llm`	LLM semantic analysis (requires API key)
`--enable-meta`	False positive filtering
`--fail-on-severity high`	Exit non-zero if HIGH/CRITICAL found (CI/CD)
`--format markdown`	Markdown report
`--format html`	Interactive HTML report
`--detailed`	Include per-finding code snippets
`--lenient`	Tolerate malformed skills

Workflow: Before Publishing a Skill to ClawHub

Run: bash scripts/scan.sh /path/to/skill
If SAFE (0 findings) → proceed to publish
If INFO only → add missing fields (license, homepage, deps) and re-scan
If MEDIUM+ → investigate and fix before publishing

Notes

"No findings" does not guarantee a skill is 100% safe -- it means no known patterns were detected
The scanner flags the lossless-claw plugin as a false positive (file read + network send is the LCM summarization pipeline -- it is safe)
Always run at minimum --use-behavioral for dataflow coverage