The audit
The AI audit, built for your next security review.
Three to four weeks. A fixed fee. A report your buyer’s reviewer can read, mapped to the AI questions they’re already asking.
Who it’s for
You sell an AI product to larger or regulated organizations. A deal is in security or procurement review, or you know one will be. You have SOC 2 or are working on it, and it doesn’t answer the AI questions. You need independent evidence that your AI does what you say, and fails safely when it doesn’t.
What you get
Each deliverable, and what it tells your buyer (and you).
- Catastrophic-failure gate
- Pass or fail on each failure you defined as unacceptable, such as your AI showing one customer another customer’s data.
- Targeted red-team tests in your own setup
- Prompt injection, including instructions hidden in content your AI reads; jailbreak attempts against your own guardrails; leaks of your system prompt or data; your AI showing one customer another customer’s data; and agent actions that add up to harm. Tested in a test copy of your system against its job, not with a generic scanner.
- Failure score
- How often it produces an unacceptable output, by failure class.
- Capability score
- How well your AI does the job it’s for, measured on your own tasks.
- Model cost comparison
- Your current model against cheaper alternatives, on cost per acceptable outcome.
- Questionnaire mapping
- Findings mapped to the AI questions in the security questionnaire your buyer sends you, the CSA AI Controls Matrix and the OWASP Top 10 for LLM Applications.
- AIUC-1 readiness check
- Every AIUC-1 requirement that applies to your system, with our test evidence mapped to the ones testing can show, and the likely gaps. Where a requirement needs evidence our tests don’t produce, such as an external pen test of your AI endpoints (B004) or database tenant isolation (A005), the check lists it as a gap and says who usually supplies it. A readiness check, not a certification. Our free AIUC-1 checklist lists every requirement.
- Policy gap list
- The policies and processes AIUC-1 expects that you don’t have yet, as a checklist your team can work through.
- Buyer summary
- A short letter-style summary you can attach to a questionnaire response.
- Evaluation harness
- The test set and harness stay ready for monthly testing and re-audits.
How the audit runs
A single AI workflow takes about three weeks; a tool-using agent about four. Before we start: a test copy of your system with test data, API access and enough model credit. The schedule starts when it works.
- 01Free · 30 minutes
Scoping call
What your AI is for, who’s reviewing you, and whether an audit is the right fit. At signing, we send the list of documents AIUC-1 expects so you can share what you have.
- 02Week 1
Define the bar
With your team, we agree what good output looks like and which failures are unacceptable or catastrophic.
- 03Weeks 1 to 3
Build, run and review
We build a test set around your system, including targeted attempts at each catastrophic failure, then run your model and several alternatives through it, with blind grading, in a test copy of your system (your staging system or an isolated copy of it) with test data, never production. We read the code on your AI’s path (how prompts are built, what its tools can reach, where its output is stored) so each failure comes with how it is reached and what to fix. This isn’t a code security review. We also review your configuration and documents against every AIUC-1 requirement.
- 04Weeks 3 to 4
Report and readout
You get the report, the buyer summary and a call to walk through it. We’ll join a call with your buyer’s reviewer if that helps.
Pricing
Fixed fees, published. Audits are paid 50% at signing and 50% on delivery; monthly testing is billed monthly, or yearly at 15% off.
$20,000
Fixed fee
A chat, RAG, document or decision workflow. About three weeks.
$27,000
Fixed fee
An agent that takes actions, calls tools or APIs. About four weeks. Same terms.
40%
Of the audit fee
When you move to a new model or make a major change.
Optional: a broad automated attack run, $5,000 with a workflow audit, $7,500 with an agent audit. Sold only with an audit.
For comparison, a first-year SOC 2 Type 2 program costs $30,000 to $80,000, and a pen test of a mid-size SaaS app $9,000 to $20,000. Neither usually tests your AI.
Sources: Scrut, SOC 2 cost; Blaze Information Security, pen test cost.
Monthly testing
Every month we test your AI again in the test copy of your system, adding new attack and failure techniques as they emerge from our own research and the industry’s, so it is not a re-run of the same tests. We tell you as soon as something fails, and you get a signed quarterly report for your buyers.
Why every month? You don’t want to wait three months to find out your agent has been leaking information. AIUC-1 asks for outside testing at least every three months; we test every month.
$2,500
A month
6-month minimum. Paid annually, 15% off: $25,500 a year.
$3,750
A month
6-month minimum. Paid annually, 15% off: $38,250 a year.
Design partners
Be one of our first three.
We’re taking three design partners. We’ll agree a start date with each one.
- You get: the full audit, report, AIUC-1 readiness check and buyer summary for $12,000 instead of $20,000 or $27,000.
- We ask: a published case study written with you (you approve every word, and it can be anonymized) and a reference call with up to three prospective clients.
- Good fit: AI in production, an active or upcoming enterprise deal, and a team that can give us a few hours a week during the audit.
Questions
Is this a certification?
No. It’s an independent audit with evidence your buyer can review, and Autogeny is not an accredited AIUC-1 auditor. The readiness check maps our test evidence to AIUC-1 requirements and shows likely gaps. We supply the testing evidence and a list of the policies you still need. We don’t write those policies, because we would then be auditing our own work. We can introduce you to a partner who does. The report alone doesn’t certify you.
We already have SOC 2. Why do we need this?
SOC 2 covers your infrastructure and processes. It doesn’t test whether your AI gives wrong answers, shows one customer another customer’s data or follows hidden instructions. That’s what buyers’ AI questions are about.
Do you need our source code?
We need read access to the code on your AI’s path: how prompts are built, what its tools can reach and where its output is stored. That isn’t a code security review. We don’t need production access: we test your system as it runs on a test copy (your staging system or an isolated copy of it) with test data. We’ll agree access during scoping, and work under an NDA and a services agreement.
What happens to our data?
We work on a test copy of your system (your staging system or an isolated copy of it) with test data, never production, and read the code on your AI’s path. Your data stays in isolated test environments for the audit and is deleted afterward unless you start monthly testing. We never use your data to train models or to build tests for other clients.
What if we fail?
You find out before your buyer does. The report says what failed and what would fix it, whether that’s a different model or a safeguard in your own system.
Do you sell models or build AI systems?
No. We don’t sell models, guardrails, runtime protection or a testing platform, and we take no referral fees, so nothing we find turns into something we sell you.
Does the audit include a red team or code review?
It includes a targeted red team: adversarial tests built around your AI’s job and run inside a test copy of your system, covering prompt injection, jailbreak attempts against your guardrails, data leaks and harmful agent actions. A broad automated attack run (large open-source attack libraries, with every hit reproduced before we report it) is an optional add-on, sold only with an audit: $5,000 with a workflow audit, $7,500 with an agent audit. We don’t do code security reviews or pen tests of your app; we can point you to independent firms, and we take no referral fee.
Send us the questionnaire.
If a buyer has sent you an AI security questionnaire, bring it to a scoping call. We’ll tell you which questions the audit answers and which it doesn’t.