Security audit
Security checks for vulnerabilities and agentic risk
This skill provides a coherent product-review workflow with local document generation and manageable privacy and recommendation-quality caveats.
Run scans in the intended project and review generated judgments before relying on or committing them. For the meeting example, obtain participant consent and use approved processing tools or redacted samples. Adopt the CLAUDE.md template only if you want its ongoing feature-development gate.
Without declared permissions the skill's intent is opaque and cannot be validated.
The recommended experiment uses real meeting recordings with transcription tools and an LLM without requiring participant consent or approval of the processing destination. Although the example recognizes recording acceptance elsewhere, that does not authorize sharing confidential meeting content with external providers; following this workflow could expose personal or proprietary information.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
pass
try:
log = subprocess.check_output(
["git", "log", "--oneline", "-10"],
cwd=root, stderr=subprocess.DEVNULL
).decode().strip().split("\n")
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
).decode().strip().split("\n")
summary["recent_commits"] = [l for l in log if l]
first = subprocess.check_output(
["git", "log", "--oneline", "--reverse", "-1", "--format=%ar"],
cwd=root, stderr=subprocess.DEVNULL
).decode().strip()
The script describes itself as a collector rather than a substantive judge, yet computes verdicts from coarse rules and publishes them as final judgments without a provisional label. Because the resulting document is intended to govern implementation decisions, users or agents could mistake an unreviewed heuristic result for an evidence-based assessment, incorrectly blocking or greenlighting work.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
- You're about to build something because a competitor has it
- A PR adds scope beyond what JUDGMENT.md describes
**DO NOT trigger the gate — proceed without checking:**
- Bug fixes (behavior unchanged)
- Performance optimizations on existing features
- Refactors that don't change user-facing behavior
The instruction requires every verdict to include both English and Chinese, regardless of the user's language preference. No opt-in or justification for this mandatory language constraint is provided.