Focus
App + process
Product features and how engineers use AI
AI security · LLM apps · Data governance
Security audits for AI-powered products and workflows
I review how your product and engineering workflows use AI — prompts, data flows, API keys, RAG pipelines, and user-facing features — so you ship innovation without preventable leaks or compliance surprises.
Overview
AI features introduce new attack surfaces: prompt injection, over-permissive tool use, training data in logs, customer PII in embeddings, and API keys in repos or browser bundles.
I audit AI projects with a builder’s mindset — I run production AI monitoring (Observinio) and ship LLM-integrated products — so findings are specific and fixable, not generic security theater.
Coverage includes application architecture, third-party model usage, vibe coding risks in your repo, and policies your team can actually follow.
At a glance
Standards
OWASP LLM
Mapped to practical mitigations
Output
Risk register
Severity, owner, and remediation steps
What you get
Threat model for AI features
User inputs, tool calling, file uploads, and admin surfaces that can be abused.
Data flow and retention review
What leaves your boundary, what providers log, and minimization opportunities.
Secrets and supply chain check
API keys, MCP servers, dependencies, and CI/CD exposure from AI tooling.
Governance recommendations
Policies, logging, red-team scenarios, and incident response hooks.
AI security finding categories
Relative frequency in audits of early-to-mid-stage AI products.
How it works
Week 1Architecture map — Models, vendors, data stores, and user roles.
Week 1Testing pass — Prompt injection probes and permission edge cases.
Week 2Engineering review — Repos, env handling, and vibe coding config.
Week 2Report & briefing — Risk-ranked findings with fix guidance.
Outcomes teams care about
- Clear picture of AI-specific risks before enterprise sales or compliance reviews
- Actionable fixes engineering can schedule
- Confidence for founders explaining safety to customers and investors
- Alignment between security, legal, and product on acceptable use
Frequently asked questions
Is this a formal penetration test?
It is a focused AI security assessment, not a full pentest certification. I can partner with specialized firms if you need formal compliance attestation.
Which AI stacks do you cover?
OpenAI, Anthropic, OpenRouter, RAG with vector DBs, agent frameworks, and custom MCP integrations — plus how your team uses Cursor or Claude Code on the codebase.
Do you help remediate issues?
Yes. Implementation support for guardrails, logging, and secure architecture is available after the audit.
We are pre-launch — is it too early?
Pre-launch is often the cheapest time to fix data handling and permission models before real customer data flows through the system.
Shipping AI features responsibly?
Describe your AI product, data sensitivity, and launch timeline. I will outline audit scope and deliverables.