Audit

Security Auditor — Intelligence Layer for Sentinel

Sentinel runs the tools. This skill provides the brain.

When Sentinel output is available: enriches every SENTINEL-XXX finding with attack chains, exploitability assessment, fix code, and false-positive analysis.

Without Sentinel output: performs full static security analysis on any code, IaC, or config using OWASP Top 10 reasoning, logic vulnerability detection, and the reference files below.

Reference files — load as needed

• references/owasp-top10.md — load when analyzing application code (injection, auth, XSS, etc.)
• references/iac-checklist.md — load when reviewing Dockerfile, k8s YAML, Terraform, GitHub Actions, nginx
• references/crypto-guidance.md — load when crypto issues arise (passwords, JWT, TLS, random, AES)

Only load the reference file relevant to the current analysis. All three together = ~9KB context.

---

Input modes

Mode	Trigger	What to do
Sentinel companion	User pastes consolidated.json or runs /sentinel:audit after a scan	Enrich each SENTINEL-XXX finding
Code audit	User pastes source code or uploads a file	Full OWASP analysis, 6 layers
IaC audit	Dockerfile, k8s, Terraform, CI YAML	Load iac-checklist.md, produce hardened version
Focused question	"Is this JWT safe?" / "Can this be injected?"	Direct answer, no full report
Threat model	Architecture description, API spec, data flow	Assets, trust boundaries, attack vectors

---

Sentinel companion mode (primary use case)

When the user provides Sentinel's consolidated.json, process each finding:

## Enhanced Analysis: SENTINEL-001
Tool: semgrep | Severity: CRITICAL | CWE: CWE-89

Confirmed? YES / LIKELY / FALSE POSITIVE
[Reasoning from code context]

Exploitability: Trivial / Easy / Moderate / Hard / Theoretical
[Why: "Endpoint is public, no auth required, input reaches sink in 2 hops"]

Attack chain:
1. Attacker sends: GET /users?id=1 OR 1=1--
2. Input reaches db.query() at user.ts:45 without sanitization
3. Result: full users table returned — credentials, emails, PII

Fix (complete, runnable):
[corrected code with inline comments explaining each security change]

OWASP: A03:2021 Injection | Compliance: PCI-DSS 6.5.1, SOC 2 CC6.1

After all findings, add:

## Findings Sentinel missed (logic/architecture)
[Vulnerabilities tools cannot detect: business logic flaws, IDOR, race conditions]

## False positive analysis
[Sentinel findings that appear to be false positives, with reasoning]

## Combined risk score
Sentinel score: XX/100
Manual analysis delta: [+/-] because [reason]
Adjusted score: YY/100

---

Full audit — 6 layers

When given code to audit without Sentinel output:

Layer 1 — Attack surface mapping

Entry points, trust boundaries, sensitive assets, external dependencies, auth perimeter.

Layer 2 — OWASP Top 10 analysis

Load references/owasp-top10.md. For each finding:

[SEVERITY] [CWE-XXX] Title
Location: file:line | Confidence: HIGH/MEDIUM/LOW

Attack chain:
1. [specific input] → [specific function] → [missing control] → [impact]

Evidence:
  [exact vulnerable code]

Fix:
  [corrected code + one-line explanation]

Severity: CRITICAL (RCE, auth bypass) | HIGH (SQLi, XSS, SSRF) | MEDIUM (CSRF, weak crypto) | LOW (headers, info disclosure) | INFO (best practice)

Layer 3 — Logic vulnerabilities (tools miss these)

• Broken business logic: rate limit bypass, negative prices, TOCTOU, state machine skips
• Auth/authz: JWT alg:none/kid injection/weak secret, IDOR, OAuth state missing, session fixation
• Crypto misuse: ECB mode, reused IVs, timing attacks (== vs hmac.compare_digest), Math.random() for tokens
• Race conditions: check-then-act, double-spend, insufficient locking

Layer 4 — IaC & infrastructure

Load references/iac-checklist.md when Dockerfile, k8s, Terraform, or CI/CD files are present.

Layer 5 — Secrets & credentials

API keys, private keys, DB connection strings, base64-encoded secrets, credentials in comments or URLs. For each: identify service, assess if real vs placeholder, recommend rotation + secrets manager migration.

Layer 6 — Dependencies (pattern-based, no live CVE lookup)

Known vulnerable package patterns, typosquatting risk, unpinned versions. Note: for live CVE data → run /sentinel:sentinel quick (SCA scanner).

---

Output format

# Security Audit Report
Component: [name] | Date: [today] | Scope: [what was reviewed]
Summary: X critical, Y high, Z medium, W low

Risk Score: XX/100

## Attack Surface
[Layer 1 output]

## Findings
[One block per finding, format above]

## What's secure ✓
[2–4 specific things done well — always include]

## Priority fix order
1. [Fix today — why]
2. [Next sprint — why]
3. [Nice to have — why]

## Sentinel integration note
[If Sentinel ran: "N findings confirmed, M appear to be false positives"]
[If Sentinel not installed: "Run /sentinel:sentinel for dependency CVEs and git history secrets scan"]

---

Fix code standards

Every fix must be: runnable (not pseudocode), idiomatic (language best practices), explained (one comment per change), non-breaking (preserves functionality).

# BEFORE (vulnerable):
query = f"SELECT * FROM users WHERE email = '{email}'"

# AFTER — parameterized query separates code from data.
# DB driver handles escaping; user input never reaches the SQL parser.
query = "SELECT * FROM users WHERE email = %s"
cursor.execute(query, (email,))

---

Scope honesty — always be explicit

• "This covers the auth module only. Payment processing was not reviewed."
• "Cannot confirm SSRF without seeing validateUrl() — not provided."
• "SENTINEL-042 appears to be a false positive: escape() at line 18 sanitizes before the sink at line 34."
• "For live CVE data, this analysis cannot substitute for Sentinel's SCA scanner."

Never fabricate CVE IDs. Describe the vulnerability pattern without asserting a specific CVE number if uncertain.

---

Report Format

Format your final output following the standard Sentinel report structure defined in ${CLAUDE_SKILL_DIR}/../../templates/report.md. Use your skill's domain-specific finding IDs (e.g. STRIDE-SPOOF-001, RT-SK-001, API-001) in the Finding ID column. Include the Security Scorecard and Findings sections as a minimum. Omit the Cross-Validation Summary section if you ran only AI analysis (no tool comparison).