CAIN-42 CAIN StudioFixed-price assessment · 3 business days
Focused AI Agent / MCP Security Assessment
One agent or one MCP server, up to 20 tools. We find what an unauthorized call could do, write the default-deny policy that stops it, and give you a signed bundle your buyer can verify without trusting us.
Prefer to talk first? Email vincent@firmeai.com. Payment is not a waiver of the scope: we confirm the scope in writing within one business day, and if it does not fit we refund in full before any work starts.
Who this is for
A small team about to give an agent tools that touch files, shells, APIs or customer data — and who cannot currently show a buyer, insurer or security reviewer what stops a bad call.
Typical trigger: “Our pilot customer asked how we prevent the agent from deleting data, and we have no answer that is not a paragraph in a doc.”
What you get
1. Findings report
Every finding ranked by severity, with the tool, the argument pattern and the evidence. Heuristic candidates are labeled as candidates, not facts.
2. Default-deny policy
A policy.json you can load into MCPGate: allow, approval, deny, with wildcards only where you ask for them.
3. Signed evidence bundle
Ed25519-signed, SHA-256-chained, with verify_keystone.py that imports no CAIN code and reports every problem it finds.
4. Walkthrough
A 30-minute call (recorded on request) covering the findings, the policy and exactly how to re-run verification.
Scope
Included
- One agent or one MCP server, up to 20 tool definitions.
- Tool-surface review: injection phrases, secret exposure, unconstrained exec/SQL/path, over-broad permissions, destructive verbs.
- One policy recommendation with per-tool verdicts.
- One evidence bundle and one verification walkthrough.
Not included
- Live exploitation, penetration testing, or scanning your production network.
- Code changes (available as an add-on at $500–$2,500 depending on scope).
- Compliance certification, legal advice, or an audit opinion.
- More than one revision of the report.
What we need from you
- A JSON export of the tool definitions (name, description, input schema), or a read-only staging endpoint.
- The policy you think you have, if one exists.
- One named technical contact for questions.
- Written confirmation that you are authorized to share the material.
If the export contains secrets, redact values and keep the shape. We do not need live credentials.
Timeline
| When | What |
|---|---|
| Day 0 | Scope confirmed; payment received; intake returned. |
| Day 1 | Tool surface scanned; candidates reviewed by a person. |
| Day 2 | Policy written; bundle generated; self-verification run. |
| Day 3 | Delivered with walkthrough. You verify independently. |
If we cannot produce a bundle that passes its own verifier, we do not send it and you do not pay the balance. For this offer payment is in full, so the remedy is a refund of the fee.
Questions a careful engineer asks
Why buy this?
You want a written, per-tool answer to "what can an agent do through this server?" before a customer or reviewer asks. If you already have that, you probably do not need us.
What does the signature prove?
That the report, findings and policy are the ones we issued and have not changed since. It does not prove your system is secure, that we found everything, or that a third party checked our work.
Who reviews the findings?
A FIRME AI CORPORATION engineer, named in the signed report, confirms or drops every engine finding. The engine is deterministic and heuristic. This is not an independent audit.
How is my source code protected?
We do not need it. The review reads tool names, descriptions and input schemas. Do not send source, credentials or production data unless we agree that in writing first.
What happens to my data?
Used only for this engagement, kept in an access-controlled directory, deleted 90 days after delivery unless you ask otherwise in writing. Not used to train a model. Email is not encrypted by us, so agree on a channel first if your policy needs it. Intake guide.
What if you find something serious?
We tell your named technical contact privately within one business day of confirming it, without waiting for the report. We do not publish it, tell anyone else, or test it live. Fixing it is your call.
What happens after I pay?
Stripe emails a receipt. You send the intake. We confirm scope in writing within one business day, or refund you in full. The 3-business-day clock starts when your inputs are accepted.
What can it miss?
Anything outside the tool definitions: authentication, transport, token storage and server logic. Those need a code review or a penetration test, and we will say so rather than imply we covered them.
What this is not
It is not a certification. It is not a guarantee that the agent cannot be abused. The scanner is heuristic and reviews definitions and inputs you provide, not a running production system. A signed bundle proves the record is intact and bound to the report — not that the system is secure or that a third party reviewed it. Findings may be incomplete. We will say so in the report rather than padding it.
Request it
Two ways to start. Pay now and send the intake within one business day, or send the intake first and we reply with the scope and the link. Either way, work starts only after the scope is confirmed and the payment has cleared.
Prefer email? vincent@firmeai.com. Payment is handled by Stripe; we never see or store your card number.