The 2026 AI Code Audit Checklist: What Every CTO Needs to Review

The Checklist Your AI Assistant Will Never Give You

Every engineering team using AI code assistants needs a systematic way to verify what the AI produced. This is the checklist we use internally at Sherlock Forensics when auditing AI-generated codebases. It is organized by security category with specific items to check, what AI typically gets wrong and severity ratings.

Print it. Bookmark it. Run it against every AI-generated codebase before it reaches production. If you built with Cursor, Bolt or Lovable, also review our vibe coding security audit service.

1. Dependency Verification

Check	What AI Gets Wrong	Severity
Verify every import exists on its registry	Hallucinated packages that attackers register with malware	Critical
Check all dependencies against NVD for known CVEs	Uses outdated package versions with known vulnerabilities	High
Verify package name spelling (typosquatting)	Imports lodahs instead of lodash	Critical
Map transitive dependency tree	Does not consider indirect dependencies at all	Medium

2. Secrets Scanning

Check	What AI Gets Wrong	Severity
Scan for hardcoded API keys and credentials	Embeds sk-proj-, AKIA and JWT secrets from training data	Critical
Run entropy analysis on all string literals	High-entropy strings that are real credentials disguised as placeholders	Critical
Check git history for committed-then-removed secrets	Commits secrets then "fixes" by removing them (still in history)	High
Verify environment variable usage for all credentials	Uses process.env in some files but hardcodes in others	High

3. Authentication Flow Review

Check	What AI Gets Wrong	Severity
Rate limiting on login and password reset	No rate limiting on any endpoint. Brute-force is trivial.	High
Token generation uses cryptographic randomness	Uses Math.random() or Python random for security tokens	Critical
JWT tokens have reasonable expiration	Tokens set to expire in 365 days or never	High
Password reset tokens are single-use	Tokens remain valid after use	High
Session IDs rotate after privilege change	Same session from anonymous to admin	Medium

4. API Security

Check	What AI Gets Wrong	Severity
Object-level authorization on every endpoint	Checks authentication but not authorization. User A can access User B data.	Critical
Response filtering (no excessive data exposure)	Returns entire database objects including internal fields	High
Rate limiting per endpoint and per user	No rate limiting anywhere. Entire API is scrapeable.	High

5. Input Validation

Check	What AI Gets Wrong	Severity
All database queries use parameterized statements	String concatenation in SQL. CWE-89	Critical
No shell command construction from user input	Uses exec() or system() with user data. CWE-78	Critical
File paths validated against traversal	Accepts ../../etc/passwd in file operations	High
No unsafe deserialization of user input	Uses pickle.loads(), unserialize() on untrusted data. CWE-502	Critical

6. Output Encoding

Check	What AI Gets Wrong	Severity
HTML encoding on all user-controlled output	Injects user data with innerHTML or v-html. CWE-79	High
Context-appropriate encoding (HTML, JS, URL, CSS)	Encodes for HTML but not for JavaScript or URL contexts	Medium

7. Session Management

Check	What AI Gets Wrong	Severity
Cookies set with Secure, HttpOnly, SameSite	Omits security attributes on session cookies	High
CSRF protection on state-changing requests	No CSRF tokens. Forms submit cross-origin.	High
Session invalidation on logout	Clears cookie client-side but session remains valid server-side	Medium

8. Error Handling

Check	What AI Gets Wrong	Severity
No stack traces in production responses	Returns full error objects with file paths and line numbers	Medium
Generic error messages for authentication failures	Differentiates "user not found" from "wrong password" (user enumeration)	Medium

9. Logging

Check	What AI Gets Wrong	Severity
Authentication events logged	No logging at all. Zero forensic trail after breach.	High
Authorization failures logged	Returns 403 but does not record the attempt	Medium
Logs do not contain sensitive data	Logs full request bodies including passwords and tokens	High

Use This Checklist. Then Call Us.

This checklist is a starting point. It covers the vulnerability patterns we find most often in AI-generated codebases. Running through it internally will catch the obvious issues.

The non-obvious issues require a professional auditor who has seen hundreds of AI-generated codebases and knows where the subtle vulnerabilities hide. That is what Sherlock Forensics does. We have been doing security work for over 20 years. AI code auditing is the newest application of the same investigative methodology. Read our full AI code audit service details or order a quick audit online starting at $1,500.

Reference standards used in this checklist:

• OWASP Top 10 (2021)
• OWASP API Security Top 10 (2023)
• MITRE Common Weakness Enumeration
• NIST National Vulnerability Database