pii-detector — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited pii-detector (Agent Skill) and scored it 87/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 3 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 3 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Automatic PII checks during development. Fires on both what the user asks for AND what Claude is about to generate — not just on keyword matching.
Check before generating. Never produce code and then suggest fixes. The sequence is always: detect intent → run relevant checks → generate correct code.
If there is ANY ambiguity — check it. False positives are cheap. Missed PII in production is a breach, a fine, or both.
When suggesting fixes, use the language, framework, and idioms of the code being reviewed. Never suggest a fix in a different language than the one being written. If the codebase is Python, fix in Python. If Rails, fix in Ruby. If Go, fix in Go. Claude infers the language from context — no need to specify.
Read the user's request AND what Claude is about to produce. Choose the mode first — this determines everything else.
Planning / Review mode — user is discussing, designing, reviewing, or asking Claude to read/check existing code. No new code being generated. → Load: modes/planning.md → Enrich Claude's natural response with PII notes. No standalone report.
Signals: "review", "read", "look at", "check this", "share thoughts", "how should I build", "I want to build X", "help me design", "feedback on", "thoughts on", reading or analyzing an existing file without generating new code.
Generation mode — Claude is about to write new code from scratch. → Load relevant layer(s) from the table below. Check first, generate second.
Signals: "build it", "create", "generate", "write", "implement", "scaffold", "make", "add this feature" — active code production.
Repo scan mode — user explicitly requests a full audit. → Load: modes/repo-scan.md + all layers.
Signals: "scan my repo", "full audit", "PII report", "check all my models", "audit my codebase".
When in doubt between planning/review and generation: default to planning/review. Enriching a response is always safer than running a check that wasn't needed.
Only applies when generating new code. Multiple layers can apply simultaneously.
| What's being built | Layers to load |
|---|---|
| Database model / schema / migration | patterns/fields.md + rules/non-negotiables.md + modes/inline.md |
| Login / signup / auth / registration | layers/auth-sessions.md + rules/non-negotiables.md |
| Frontend form / page / component | layers/frontend.md |
| API route / controller / serializer | layers/api-layer.md |
| Middleware / request handler / logger | layers/data-in-transit.md |
| JWT / token / session / cookie code | layers/auth-sessions.md |
| Webhook handler | layers/data-in-transit.md + layers/api-layer.md |
| Seed file / fixture / factory / test helper | layers/testing-seeding.md |
| Delete / purge / anonymize / export function | layers/data-lifecycle.md |
| Cron job / background worker / cleanup task | layers/data-lifecycle.md |
| External API client / HTTP call | layers/data-in-transit.md |
| Analytics / tracking integration | rules/leakage-vectors.md |
| Error handling / monitoring setup | rules/leakage-vectors.md |
| Full repo / codebase submitted | ALL layers — load modes/repo-scan.md |
When multiple layers apply (e.g. building an auth API endpoint): load all relevant layers and merge the checks.
Always load patterns/fields.md when a model or schema is involved. Always load rules/non-negotiables.md when any PII field is detected.
| Regulation | What It Means for Code |
|---|---|
| CCPA/CPRA | Every PII field must be deletable on request |
| HIPAA | Health data encrypted at rest + in transit; access logged on every read |
| PCI-DSS | Never store CVV; tokenize card numbers; never log payment bodies |
| COPPA | Flag age/dob fields; under-13 = parental consent, no behavioral tracking |
| GLBA | Financial data: encrypt, access control, audit logs |
| BIPA | Biometric: written consent before storage; retention limit required |
| FERPA | Student records: strict access controls, no sharing without consent |
| FTC Act | Do exactly what your privacy policy says |
Never generate broken code first. Never suggest fixes as an afterthought. The check and the corrected output happen in a single pass.
Inline (any single-task coding):
⚡ PII Check — [what's being built]
🔴 Fix now:
[specific issue] → [exact fix]
🟡 Fix soon:
[specific issue] → [exact fix]
🟢 Consider:
[suggestion]
Regulations: [only ones actually triggered]Then generate the corrected code immediately after.
Repo scan: structured report with file + line references. Load modes/repo-scan.md.
Rules for all output:
ssn with AES-256, store as BYTEA" not "encrypt sensitive fields"~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.