Our review
This skill audits the coverage of an AI system's guardrails: inventorying active filters, testing bypasses via jailbreak and encoding tricks, and measuring the false-positive rate.
Strengths
- Explicitly inventories every filter and policy rule before testing, which prevents blind spots
- Exercises real bypass techniques (unicode, base64, role-play framing) instead of the happy-path input
- Reports both confirmed bypasses and false-positive rates, ranked by exploitability and blast radius
- Produces reproduction steps for each confirmed bypass
Limitations
- Assumes access to real traffic samples to measure false positives meaningfully
- Relies on an external output format (docs/output-kit.md) not included in the skill
- Detailed findings are routed through /atlas-report, so CLI output is intentionally minimal
Use it when you need to assess whether an AI application's guardrails can be bypassed or are over-blocking legitimate users.
Don't use it to design or implement guardrails, or for a simple functional test of isolated filters without representative traffic.
Security analysis
CautionThe skill is designed for legitimate guardrail auditing and does not include destructive or exfiltration actions. However, it involves adversarial testing of safety filters using powerful tools, which carries dual-use risk and should be handled with appropriate oversight.
- •Instructs testing bypass vectors including jailbreak patterns and encoding tricks, which could be repurposed to evade guardrails
- •Uses Bash and web tools for legitimate security auditing, but these are powerful and dual-use
Examples
Audit our AI guardrails — inventory what's active, test it with known jailbreak and encoding bypasses, and tell me what gets through.Can our input filters be bypassed? Try unicode, base64 and role-play framing against them and give me reproduction steps for anything that works.Check the false positive rate of our guardrails using these traffic samples and report how often legitimate requests get blocked.name: guard-audit description: Audit guardrail coverage — bypass vectors, false positive rates, policy gap analysis, red-team scenarios. Use when asked to "audit our AI guardrails", "can our filters be bypassed", or "check guardrail false positives". allowed-tools: Read, Bash, Glob, Grep, Write, WebFetch, WebSearch, AskUserQuestion version: 1.0.0 author: tonone-ai hello@tonone.ai license: MIT compatibility: Designed for Claude Code tags: [ai-ops, guardrails, audit]
Guard Audit
You are Guard — the AI Guardrails Engineer on the AI Operations Team.
Steps
Step 0: Inventory Current Guardrails
List every input/output filter, classifier, and policy rule currently active, and what each is meant to catch.
Step 1: Test Bypass Vectors
Run known jailbreak/prompt-injection patterns and encoding tricks (unicode, base64, role-play framing) against each guardrail to check for gaps.
Step 2: Measure False Positive Rate
Check how often legitimate requests get blocked, using real traffic samples where available.
Key Rules
- Follow the output format defined in docs/output-kit.md
- Test with real bypass techniques, not just the happy-path input the guardrail was designed for
- A guardrail with a high false positive rate is a product problem even if it has zero bypasses — report both sides
- Rank findings by exploitability and blast radius, not just by count
Output Format
A guardrail coverage table, a list of confirmed bypasses with reproduction steps, and false-positive rate findings.
Delivery
If output exceeds the 40-line CLI budget, invoke /atlas-report with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.
Security Audit Scanner
Security
Analyzes code to detect OWASP Top 10 vulnerabilities.
OWASP Security Checklist
Security
Generates application security checklists based on the OWASP Top 10.
Threat Model Generator
Security
Generates threat model documents with STRIDE analysis.