▸ What we offer
AI agent MCP security audit $500 · focused single-server audit $300
MCP is the plug-in system that gives an AI assistant its tools — file access, databases, APIs, whatever you wired up. So: you deployed an AI agent with MCP tool access. Can it run DDL on your production database? Export customer data? Escalate its own privileges? If you cannot answer from evidence, that is the audit.
We build MCP security tooling ourselves — a deny-by-default MCP firewall written in Rust (14 crates, 512 tests, signed releases), carried through a design-failure audit of our own code: 155 findings, 107 fixed in code with a reproducing test written first, and 48 deliberately deferred with a tracking ID and the reason — because a gap you can name is safer than a gap closed on paper.
- Full MCP configuration audit — every server your agent can reach, every tool it can call, every permission it holds.
- Ranked vulnerability assessment — a kill-chain ordered by blast radius: the assumption, the perturbation that breaks it, the attack path, the minimum structural fix.
- Hardening configuration — deny-by-default allowlists, session budgets, audit trail, drift detection for your specific setup.
- Report — every finding categorized, confidence-scored, and traced to the evidence behind it.
3–5 business days · structured report + issue-tracker export · firewall source and our own audit report available to clients on request.
Structural code review $250 · up to 10k lines · $500 to 50k · larger by agreement
Most code review is dead reckoning — “looks good to me” adds no information. We review against a 22-category structural taxonomy — a named checklist of the specific ways code goes wrong — so every finding gets a specific diagnosis: not “this is bad” but Counterfeit — a null check that does not prevent the dereference; not “this is messy” but Deceptive — the comment says O(1), the implementation is a linear scan.
- Full taxonomy pass — every file inspected against all 22 categories, in a 5-phase process: triage, inspection, judgment, verification, record.
- Categorized findings — category, confidence, location, severity, and how deeply each was verified.
- Structural recommendations — the fix for the class, not just the instance.
- Durable review record — every finding traced to evidence, so you can re-check any of it.
2–5 business days depending on size · structured review document.
AI agent system-prompt architecture $300 · multi-agent systems $500
A system prompt is not a list of instructions; it is the architecture that decides what your agent claims, when it stops, and what it does when it is wrong. Most agent prompts are wishful thinking in natural language. We design them as load-bearing structure:
- Epistemic discipline — claims carry confidence labels; the agent never says “confirmed” without proof.
- Halt protocols — tested halt words that stop the agent mid-action, not after it finishes the sentence.
- Evidence gates — a table of which evidence types must exist before which claim types are allowed.
- Chain of command — who outranks whom, when to defer, when to escalate, when to refuse.
- Test scenarios — 12+ probes per dimension so the prompt is validated, not vibes.
3–5 business days · system prompt + architecture document + test scenarios.
How an engagement works
Write to us through the contact form with what you have — a repo, an agent config, a prompt, or just the worry. We reply with a fixed scope and a fixed price before any work starts; if we are not the right fit we say so and that costs nothing. Work is delivered as documents you keep, with every finding traced to evidence. What we did not check is listed next to what we did — the same standard we hold our own products to on this site.