Security Reviewer
A pragmatic AppSec review pass for any modern stack. Output is short, severity-ranked findings with file paths, evidence, and a concrete fix — never a 30-page "report".
When to run
Run on any of:
- New HTTP route, RPC handler, or queue consumer
- Auth, session, RBAC, or tenant-isolation change
- Frontend rendering / templating change (XSS surface)
- Outbound HTTP, file fetch, or URL handling (SSRF surface)
- Database query change (injection / IDOR surface)
- Dependency manifest change or lockfile bump
- Config / IaC change (CORS, headers, CSP, IAM, secrets)
- End of a major feature, pre-release, or before opening a PR labelled
security-sensitive
If unsure, run it. False positives are cheap; missed CVEs are not.
Eight-point check (language-agnostic)
These eight categories drive every review. Every finding maps to one.
- Authn/Authz — every route has explicit auth + role/tenant checks; no implicit "logged-in means allowed"
- Output encoding — all user-influenced HTML/markdown/JSON/log output is encoded for its sink
- SSRF & outbound HTTP — URL allow-listed, redirects bounded, IMDS / link-local / RFC1918 blocked unless intentional
- Secrets & key material — no client-side storage, no hardcoded keys, no secrets in URLs or logs
- Input handling — parameterised queries, path containment, redirect target validation, deserialisation guards
- Config & headers — CSP, HSTS, X-Content-Type-Options, secure cookies, CORS allowlist, rate limits, dev-bypass off in prod
- Supply chain — dependencies pinned, lockfile committed, manifest scanned, no unknown registries, install scripts reviewed
- Failure modes — fail-closed on auth/policy errors, no broad
except/catch, errors don't leak stack traces or PII
Routing
When language is known, load the matching reference and use its diagnostic patterns. When mixed, run all relevant ones.
| Detected from… |
Read… |
pyproject.toml, requirements*.txt, Pipfile, setup.py |
languages/python.md |
package.json, *.ts, *.tsx, *.js, *.jsx |
languages/javascript-typescript.md |
go.mod |
languages/go.md |
Cargo.toml |
languages/rust.md |
pom.xml, build.gradle*, *.kt, *.java |
languages/java.md |
Gemfile, *.rb |
languages/ruby.md |
*.csproj, *.sln, *.cs |
languages/dotnet.md |
Then consult the cross-cutting checklists:
checklists/owasp-asvs.md — OWASP Top 10 (2021) and ASVS L1 controls keyed to the eight-point checkchecklists/endpoint-checklist.md — per-endpoint review templateplaybooks/triage.md — what to do once a finding fires (severity rubric, fix templates, follow-up tickets)
Output format
Every review emits a finding list. Nothing else.
[SEVERITY] <one-line summary>
File: <path>:<line>
Category: <one of the 8 categories>
Evidence: <code snippet or pattern matched>
Fix: <concrete change, ideally a diff>
Refs: <CWE-XXX, OWASP A0X:2021, ASVS V-X.Y.Z>
Severities: CRITICAL (auth bypass, RCE, secret leak in prod), HIGH (data exposure, SSRF to internal, missing CSRF on state change), MEDIUM (weak crypto choice, missing rate limit, stack trace leak), LOW (defence-in-depth, hygiene).
If nothing fires, say so explicitly and list residual risks or untested areas — silence is not a clean bill.
Hook integration
Three hooks ship with this skill (/.claude/hooks/):
session-start.sh — fingerprints the project on session open, runs a fast dependency audit if tooling is present, and injects a security-context block so every agent in the session starts informed.pre-bash-package-guard.sh — gates package-manager invocations (npm/yarn/pnpm/pip/uv/poetry/cargo/go get/gem/bundle/composer/dotnet add). Blocks installs of typosquatted, brand-new, or known-malicious packages; prompts for confirmation on borderline cases.post-edit-quickscan.sh — runs a fast pattern scan on files touched by Edit/Write/MultiEdit and feeds findings back to the agent that made the change.
Subagents that consume this skill should treat the ## SECURITY CONTEXT block (when injected by session-start.sh) as authoritative project metadata.
Author / provenance
Maintained by Phoenix Security as part of the Phoenix developer-tooling kit. Pairs with phoenix-final-gate (spec-pipeline review) and opengrep-rule-generator (rule synthesis from findings).
1---2name: security-reviewer3description: Multi-language security review for web apps, APIs, and CLIs. Use whenever code is added or modified — new endpoints, auth/RBAC changes, template or DOM rendering, outbound HTTP, dependency updates, IaC/config changes, or end-of-feature reviews. Covers Python, JavaScript, TypeScript, Go, Rust, Java/Kotlin, Ruby, C#/.NET, plus shared OWASP Top 10 and ASVS L1 logic. Triggers on phrases like "security review", "review this for security", "audit this code", "is this safe", "AppSec check", "threat-model this change", "supply chain risk", "before I merge", or whenever a senior engineer would pause to ask "what could go wrong here". Pairs with the `security-reviewer` subagent and the session-start / pre-bash / post-edit hooks shipped alongside it.4---56# Security Reviewer78A pragmatic AppSec review pass for any modern stack. Output is short, severity-ranked findings with file paths, evidence, and a concrete fix — never a 30-page "report".910## When to run1112Run on any of:1314- New HTTP route, RPC handler, or queue consumer15- Auth, session, RBAC, or tenant-isolation change16- Frontend rendering / templating change (XSS surface)17- Outbound HTTP, file fetch, or URL handling (SSRF surface)18- Database query change (injection / IDOR surface)19- Dependency manifest change or lockfile bump20- Config / IaC change (CORS, headers, CSP, IAM, secrets)21- End of a major feature, pre-release, or before opening a PR labelled `security-sensitive`2223If unsure, run it. False positives are cheap; missed CVEs are not.2425## Eight-point check (language-agnostic)2627These eight categories drive every review. Every finding maps to one.28291. **Authn/Authz** — every route has explicit auth + role/tenant checks; no implicit "logged-in means allowed"302. **Output encoding** — all user-influenced HTML/markdown/JSON/log output is encoded for its sink313. **SSRF & outbound HTTP** — URL allow-listed, redirects bounded, IMDS / link-local / RFC1918 blocked unless intentional324. **Secrets & key material** — no client-side storage, no hardcoded keys, no secrets in URLs or logs335. **Input handling** — parameterised queries, path containment, redirect target validation, deserialisation guards346. **Config & headers** — CSP, HSTS, X-Content-Type-Options, secure cookies, CORS allowlist, rate limits, dev-bypass off in prod357. **Supply chain** — dependencies pinned, lockfile committed, manifest scanned, no unknown registries, install scripts reviewed368. **Failure modes** — fail-closed on auth/policy errors, no broad `except`/`catch`, errors don't leak stack traces or PII3738## Routing3940When language is known, load the matching reference and use its diagnostic patterns. When mixed, run all relevant ones.4142| Detected from… | Read… |43|---|---|44| `pyproject.toml`, `requirements*.txt`, `Pipfile`, `setup.py` | `languages/python.md` |45| `package.json`, `*.ts`, `*.tsx`, `*.js`, `*.jsx` | `languages/javascript-typescript.md` |46| `go.mod` | `languages/go.md` |47| `Cargo.toml` | `languages/rust.md` |48| `pom.xml`, `build.gradle*`, `*.kt`, `*.java` | `languages/java.md` |49| `Gemfile`, `*.rb` | `languages/ruby.md` |50| `*.csproj`, `*.sln`, `*.cs` | `languages/dotnet.md` |5152Then consult the cross-cutting checklists:5354- `checklists/owasp-asvs.md` — OWASP Top 10 (2021) and ASVS L1 controls keyed to the eight-point check55- `checklists/endpoint-checklist.md` — per-endpoint review template56- `playbooks/triage.md` — what to do once a finding fires (severity rubric, fix templates, follow-up tickets)5758## Output format5960Every review emits a finding list. Nothing else.6162```63[SEVERITY] <one-line summary>64File: <path>:<line>65Category: <one of the 8 categories>66Evidence: <code snippet or pattern matched>67Fix: <concrete change, ideally a diff>68Refs: <CWE-XXX, OWASP A0X:2021, ASVS V-X.Y.Z>69```7071Severities: `CRITICAL` (auth bypass, RCE, secret leak in prod), `HIGH` (data exposure, SSRF to internal, missing CSRF on state change), `MEDIUM` (weak crypto choice, missing rate limit, stack trace leak), `LOW` (defence-in-depth, hygiene).7273If nothing fires, say so explicitly and list residual risks or untested areas — silence is not a clean bill.7475## Hook integration7677Three hooks ship with this skill (`/.claude/hooks/`):7879- **`session-start.sh`** — fingerprints the project on session open, runs a fast dependency audit if tooling is present, and injects a security-context block so every agent in the session starts informed.80- **`pre-bash-package-guard.sh`** — gates package-manager invocations (`npm/yarn/pnpm/pip/uv/poetry/cargo/go get/gem/bundle/composer/dotnet add`). Blocks installs of typosquatted, brand-new, or known-malicious packages; prompts for confirmation on borderline cases.81- **`post-edit-quickscan.sh`** — runs a fast pattern scan on files touched by `Edit`/`Write`/`MultiEdit` and feeds findings back to the agent that made the change.8283Subagents that consume this skill should treat the `## SECURITY CONTEXT` block (when injected by `session-start.sh`) as authoritative project metadata.8485## Author / provenance8687Maintained by Phoenix Security as part of the Phoenix developer-tooling kit. Pairs with `phoenix-final-gate` (spec-pipeline review) and `opengrep-rule-generator` (rule synthesis from findings).
security-phoenix-demo