AI Code Security Audit
OWASP audit specific to AI-generated code and LLM endpoints
AI-generated code has characteristic security problems: exposed API keys, permissive anon roles, missing auth checks, SSRF, prompt injection in LLM endpoints, and vulnerabilities in generated dependencies. We find them before attackers do.
AI-generated code has characteristic security problems that a human dev rarely makes: exposed API keys in the client bundle, permissive Supabase anon roles, SSRF in generated API calls, prompt injection in LLM endpoints, vulnerable versions of dependencies. We audit these patterns specifically.
The express audit combines automated tools with manual code review. We run TruffleHog and GitLeaks to detect secrets in code, git history, and comments. Snyk and OSV scanner for dependency audit prioritized by CVSS. Semgrep with rules specific to AI patterns, we hunt known anti-patterns generated by Cursor, Lovable, v0. Manual review walks each API endpoint and checks auth, authorization, IDOR, SSRF, input validation, and output encoding. For RAG apps or AI features we audit LLM endpoints, testing prompt injection, jailbreak patterns, data exfiltration through output, missing rate limiting, and unbounded token spending. The threat model documents identified attack vectors and recommendations. The report has each finding with CVSS score, exploitation steps, business impact, and concrete fix. After fixes we can do a re-audit to confirm the issues are actually resolved.
- OWASP Top 10 review specific to AI patterns, exposed keys, IDOR, SSRF, prompt injection
- Secret scanning in code, git history, and comments with TruffleHog and GitLeaks
- Dependency audit (Snyk, OSV) prioritized by CVSS with recommended safe versions
- Threat model document useful for security review, SOC 2, or enterprise sales
What's included
What You Get
OWASP Top 10 review for AI code
A check specific to AI patterns: injection (including prompt injection), broken auth, exposed secrets, IDOR, SSRF, insufficient logging.
Secret and credential scanning
TruffleHog, GitLeaks, and manual review reveal API keys, OAuth secrets, DB credentials, and private keys in code, .env, git history, and even generated comments.
Dependency audit
npm audit, Snyk, OSV scanner for JS/TS dependencies. We identify CVEs in transitive deps and recommend safe versions.
Auth/authorization review
We walk through every API endpoint and check: who has access, what permission, is there IDOR, is there SSRF, does authentication pass correctly.
Prompt injection and LLM endpoint testing
If the app calls an LLM (OpenAI, Anthropic, local models), we test prompt injection, jailbreak patterns, and data exfiltration through LLM output.
Threat model write-up
We deliver a threat model document with identified attack vectors, priorities, and recommendations. Useful for security review or compliance.
Perfect for
Perfect for
- Founders taking an AI-built app to first enterprise customers and needing a security pass
- Companies with AI-generated code that must pass internal security review or SOC 2
- Teams whose AI implemented an AI feature (LLM call) and want to verify prompt injection and data leaks
- Products with an AI-built backend holding sensitive data (PII, finance, healthcare)
Our Process
How We Work
Pre-audit
You give us code and infrastructure access. We agree on scope, code only, code + DB, code + LLM endpoints, etc.
Audit (5-10 days)
Combination of automated tools (Snyk, TruffleHog, OSV) and manual review. Threat modeling. Prompt injection testing.
Report
We deliver a report with findings by CVSS score, prioritization, and concrete remediation recommendations. Threat model document.
Remediation (optional)
If you want, we fix the critical findings. Or hand the report over and you fix it yourself.
Pricing
Pricing Tailored to Your Project
Pricing by code size and audit depth.
Express audit
From €1,400
- Automated code scan
- Dependency audit
- Secret scanning
- Top findings
- 5 days
Full audit
From €3,000
- Everything in Express audit
- Manual code review
- Auth/authorization audit
- Threat model document
- 10 days
Audit + remediation
On request
- Everything in Full audit
- Critical findings fixed
- Re-audit after fixes
- Priced by findings
FAQ
Frequently Asked Questions
Prompt injection is an attack where user input overrides LLM model instructions. An attacker can force the model to reveal system prompts, other users' data, or perform unintended actions. In AI-built apps it's a common problem because vibe-coded LLM endpoints rarely have proper input sanitization.
Yes, that's our specialty. We test prompt injection, jailbreak, data exfiltration, and rate limiting on LLM endpoints. For RAG apps we also audit vector store access security.
Express audit: 5 days. Full audit: 10 days. For large apps (50k+ LOC) we can extend or split into phases.
During the audit we run into bugs, race conditions, and problematic logic, we write them up in the report. If you want a complete bug audit, we also recommend our Cursor & Claude Code cleanup service.
We always sign NDAs. For sensitive projects we work in isolated environments (separate VM, dedicated VPN) when compliance requires it.
Services
Related Services
Cursor & Claude Code Cleanup
Refactoring AI-pair-programming outputs from Cursor, Claude Code, and Windsurf.
→Lovable to Production
Migrate Lovable.dev full-stack apps to a maintainable production stack.
→v0 & Bolt Refactor
Productize v0.dev and Bolt.new prototypes with auth, payments, and proper deploy.
→Replit & Windsurf Rescue
Stabilize and rehost apps built in Replit Agent or Windsurf.
→