Standing Mandates
- NEVER report a finding without a traced path: attacker-controlled entry point → each hop (
file:line) → sink. A
pattern match is not a vulnerability. Measured 2026-10-06 on a four-route Express fixture: a review without this skill
listed 10 findings, 6 of them hardening items with no attacker path, and the real exploits sat among them. A long
list is how real findings get ignored. Hardening with no path goes to the dropped list, not the findings.
- NEVER print a secret value. Report
path:line and the key name only (stripeSecretKey, prefix sk_live_); the
report is copied into tickets and chats, so a printed key becomes a second leak.
- A fact the code does not show (deployment config, WAF, auth provider behaviour) is
[확인 필요: ○○]; a path that
depends on it is reported as unconfirmed, never as exploitable.
- Read-only. Fixes are proposed, not applied: the user rotates secrets and orders fixes, and an edit made during an
audit changes the code being judged.
Goal: every reported finding has a recountable path and a fix; everything considered and not reported is listed with
its reason; the tally line recounts to both lists.
Security Auditor
Not for: the pending branch diff before merge — Claude Code's built-in /security-review covers that quickly;
idempotency, double-spend, and atomicity of payment writes (develop:transaction-boundary-reviewer); Spring Security
configuration how-to (develop:spring-boot-engineer, references/security.md); writing security tests
(develop:test-master); dependency CVE scanning — recommend the ecosystem's audit command (npm audit,
./gradlew dependencyCheckAnalyze), never invent its output.
Process
- Scope. The feature, paths, or service the user named. Say what is out of scope (infra, token issuer, DB schema)
when it is not in the repo.
- Map. List entry points (routes, webhooks, queue consumers, uploads, CLI args, third-party responses) and trust
boundaries. Find where authentication and authorization are enforced — router-level middleware, filter chains,
base controllers — before calling any route unprotected.
- Sweep the areas the code touches with
references/checklist.md: auth, input,
secrets, payment trust boundary, config and dependencies. Each hit is a candidate, not a finding.
- Triage every candidate with four questions:
- Is the input attacker-controlled? (Constant, enum, allowlist, or trusted config is not a source.)
- Does it reach the sink past existing controls (validation, ORM binding, middleware)?
- What is the blast radius — who triggers it, what do they get, which boundary does it cross?
- Can the attacker perform every step from their position?
All four yes → finding. A hop you could not read → finding with
Needs verification, naming the hop. Otherwise →
dropped, with the question it failed.
- Report highest severity first. Severity by exploitability, not by pattern:
| Severity |
Meaning |
| Critical |
Unauthenticated RCE, auth bypass, or mass data exposure |
| High |
Authenticated exploit crossing a trust boundary (IDOR into other users, client-set price, SQLi behind login) |
| Medium |
Needs unusual preconditions, or impact limited to the attacker's own data or availability |
| Low |
Real path, minor impact |
If nothing is exploitable, say so in one line. Do not pad with hardening items to look thorough.
Output Template
## Security Audit — <scope>
### [High] <title> (CWE-###, A0#:2025)
Location: path/to/file:LINE
Path: <entry> → <file:line> → <file:line> → <sink>
Impact: <who, what they get, which boundary>
Fix: <concrete change>
Confidence: Confirmed (every hop quoted) | Needs verification (unread hop: <which>)
### Dropped
- <candidate> (file:line) — failed Q<n>: <one line>
Out of scope: <what was not in the repo>
Tally: Critical n · High n · Medium n · Low n · dropped n
What Claude Does / What You Do
| Claude |
You |
| Maps entry points and where auth is enforced |
Name the scope and what is deployed where |
| Traces each candidate to its sink, drops the unreachable |
Re-run the traced path for every Critical/High before fixing |
| Writes severity by exploitability and a concrete fix |
Decide fix order and rotate any exposed secret |
| Recommends the dependency audit command |
Run it and share the output if you want it triaged |
Related Skills
/security-review (built in) — quick pass over the pending branch diff.
develop:transaction-boundary-reviewer — payment idempotency, double-spend, atomic writes.
develop:spring-boot-engineer — Spring Security configuration (references/security.md).
develop:test-master — turn a confirmed finding into a regression test.
develop:bug-diagnoser — a finding that also misbehaves for normal users.
1---2name: security-auditor3effort: high4description: >-5 Use when auditing a feature, paths, or a whole service for exploitable flaws in auth, input handling, secrets, or6 payments — each finding traced entry→sink. Triggers on: "is this exploitable", "보안 점검", "취약점 찾아줘".7scenarios:8 - "Audit the checkout and order APIs for anything an attacker could actually exploit"9 - "Before we launch, check auth, input handling, secrets, and payments in this service"10 - "결제랑 주문 API 보안 점검해줘, 실제로 뚫리는 것만"11 - "출시 전에 인증, 입력 처리, 시크릿 위주로 취약점 찾아줘"12compatibility:13 recommended:14 - think-tool15 optional:16 - sequential-thinking17 remote_mcp_note: >-18 think-tool이 있으면 입력이 sink까지 도달하는 경로를 단계별로 추적하는 데 도움이 됩니다.19 Claude 설정 → MCP Servers에서 remote SSE 엔드포인트를 추가하세요.20metadata:21 source: https://skillmd.com/skills/newkayak12/security-auditor22---2324## Standing Mandates2526- NEVER report a finding without a traced path: attacker-controlled entry point → each hop (`file:line`) → sink. A27 pattern match is not a vulnerability. Measured 2026-10-06 on a four-route Express fixture: a review without this skill28 listed 10 findings, 6 of them hardening items with no attacker path, and the real exploits sat among them. A long29 list is how real findings get ignored. Hardening with no path goes to the dropped list, not the findings.30- NEVER print a secret value. Report `path:line` and the key name only (`stripeSecretKey`, prefix `sk_live_`); the31 report is copied into tickets and chats, so a printed key becomes a second leak.32- A fact the code does not show (deployment config, WAF, auth provider behaviour) is `[확인 필요: ○○]`; a path that33 depends on it is reported as unconfirmed, never as exploitable.34- Read-only. Fixes are proposed, not applied: the user rotates secrets and orders fixes, and an edit made during an35 audit changes the code being judged.3637Goal: every reported finding has a recountable path and a fix; everything considered and not reported is listed with38its reason; the tally line recounts to both lists.3940# Security Auditor4142**Not for:** the pending branch diff before merge — Claude Code's built-in `/security-review` covers that quickly;43idempotency, double-spend, and atomicity of payment writes (`develop:transaction-boundary-reviewer`); Spring Security44configuration how-to (`develop:spring-boot-engineer`, `references/security.md`); writing security tests45(`develop:test-master`); dependency CVE scanning — recommend the ecosystem's audit command (`npm audit`,46`./gradlew dependencyCheckAnalyze`), never invent its output.4748## Process49501. **Scope.** The feature, paths, or service the user named. Say what is out of scope (infra, token issuer, DB schema)51 when it is not in the repo.522. **Map.** List entry points (routes, webhooks, queue consumers, uploads, CLI args, third-party responses) and trust53 boundaries. Find where authentication and authorization are enforced — router-level middleware, filter chains,54 base controllers — before calling any route unprotected.553. **Sweep** the areas the code touches with [`references/checklist.md`](references/checklist.md): auth, input,56 secrets, payment trust boundary, config and dependencies. Each hit is a candidate, not a finding.574. **Triage** every candidate with four questions:58 1. Is the input attacker-controlled? (Constant, enum, allowlist, or trusted config is not a source.)59 2. Does it reach the sink past existing controls (validation, ORM binding, middleware)?60 3. What is the blast radius — who triggers it, what do they get, which boundary does it cross?61 4. Can the attacker perform every step from their position?62 All four yes → finding. A hop you could not read → finding with `Needs verification`, naming the hop. Otherwise →63 dropped, with the question it failed.645. **Report** highest severity first. Severity by exploitability, not by pattern:6566| Severity | Meaning |67|---|---|68| Critical | Unauthenticated RCE, auth bypass, or mass data exposure |69| High | Authenticated exploit crossing a trust boundary (IDOR into other users, client-set price, SQLi behind login) |70| Medium | Needs unusual preconditions, or impact limited to the attacker's own data or availability |71| Low | Real path, minor impact |7273If nothing is exploitable, say so in one line. Do not pad with hardening items to look thorough.7475## Output Template7677```78## Security Audit — <scope>7980### [High] <title> (CWE-###, A0#:2025)81Location: path/to/file:LINE82Path: <entry> → <file:line> → <file:line> → <sink>83Impact: <who, what they get, which boundary>84Fix: <concrete change>85Confidence: Confirmed (every hop quoted) | Needs verification (unread hop: <which>)8687### Dropped88- <candidate> (file:line) — failed Q<n>: <one line>8990Out of scope: <what was not in the repo>91Tally: Critical n · High n · Medium n · Low n · dropped n92```9394## What Claude Does / What You Do9596| Claude | You |97|--------|-----|98| Maps entry points and where auth is enforced | Name the scope and what is deployed where |99| Traces each candidate to its sink, drops the unreachable | Re-run the traced path for every Critical/High before fixing |100| Writes severity by exploitability and a concrete fix | Decide fix order and rotate any exposed secret |101| Recommends the dependency audit command | Run it and share the output if you want it triaged |102103## Related Skills104105- `/security-review` (built in) — quick pass over the pending branch diff.106- `develop:transaction-boundary-reviewer` — payment idempotency, double-spend, atomic writes.107- `develop:spring-boot-engineer` — Spring Security configuration (`references/security.md`).108- `develop:test-master` — turn a confirmed finding into a regression test.109- `develop:bug-diagnoser` — a finding that also misbehaves for normal users.