zhouwenjie1
|
6a5543cd1c
|
tools: add skill security scanner with LLM intent review
Static analysis + LLM-assisted security scanner for agent_skills.
Detects dangerous tool usage, prompt injection, credential extraction,
data exfiltration, and stealth instructions in skill markdown files.
Features:
- Tool risk scoring tailored to embedded device capabilities
- Regex hard gates (critical findings cannot be overridden by LLM)
- LLM intent review (--llm-review) to reduce false positives
- Baseline diff mode (--baseline) for CI incremental scanning
- 3-level disposition: pass / review / block
Usage: python3 scripts/skill_security_scan.py [--llm-review] [--ci]
Signed-off-by: zhouwenjie1 <zhouwenjie1@xiaomi.com>
|
2026-05-12 11:39:57 +08:00 |