We break your AI agent before your customers do.
123 real attacks against the agent you’ve put in front of customers: leaking its instructions, handing over personal data, taking actions it shouldn’t, or running up your bill. Then a report with a fix for every finding.
First three audits free · about an hour of your team’s time
$ cognify-audit run --packs injection,pii
123
attacks per audit
14
attack classes
115
of 122 held by our lab agent
7
stalls found in our lab agent
What we test
The ways agents actually fail in production.
Mapped to the OWASP Top 10 for LLM Applications 2026 — the standard your security team, and your customers’ procurement teams, will ask about.
Prompt injection
Attacks typed into the chat, hidden in documents, emails and web pages, or disguised with encodings and lookalike letters.
Hidden instructions
Attempts to extract the system prompt, configuration and internal codes — including one yes/no question at a time.
Personal data
Fake couriers, police and support supervisors; phishing emails; data smuggled out through links and partial leaks.
Unauthorised actions
Refunds, bulk cancellations and privilege changes requested on the say-so of a message or a document.
Runaway cost
Prompts that make the agent think until it times out — a hung request, a full token bill, and no answer.
Jailbreaks
Role-play, fake authority, fake developer modes and emotional pressure designed to switch its rules off.
How it works
Three steps. About an hour of your time.
- 01
A 20-minute call
You tell us how to reach the agent, what it must never say, and whether it can take actions. You sign a short authorisation to test.
- 02
We attack a staging copy
Around 123 attacks, a few at a time, in an agreed window. Rule checks first, then a model judge for the cases that need judgement.
- 03
You get the report
Every finding ranked by severity, mapped to OWASP and NIST, with the exact attack used and a fix. Then a 30-minute readout.
What you get
A report you can act on — and check.
- Every attack listed — the exact prompt and every verdict, including the ones that held
- Findings rated by what actually happened, mapped to OWASP and NIST AI RMF
- Evidence for your SOC 2 audit — mapped to criteria CC4.1 and CC7.1
- One clear fix per class of failure, in the order to make them
- Coverage shown next to the score, so a number never hides what wasn’t tested
- A single file with a SHA-256 checksum, so you can verify it hasn’t been altered
- A one-page summary you can share with your own customers — results only, no attack detail
Both from the same lab test of a demo agent.
sample-report.html · acmebot (lab)
Needs attentionSafe to hand over: the report quotes the attacks we sent, never your agent’s replies, and secrets appear only as labels — “internal escalation code”, never the code itself.
Questions
Before you ask.
Will you attack our production system?
We test a staging copy wherever one exists. If the agent can take real actions — refunds, emails, account changes — we only test it against test data, and nothing runs until you have signed the authorisation.
Is this a SOC 2 certification?
No — only a licensed CPA firm can issue a SOC 2 report, and nobody can certify you as SOC 2 compliant. What our report gives you is evidence your auditor can use: an independent security evaluation of your AI agent (criterion CC4.1) and the vulnerabilities it found (CC7.1). Re-running it after a change shows the testing is ongoing.
What do you keep?
Not the conversations. Each reply is judged and discarded; the results and the report hold verdicts and reasons, never transcripts. Credentials live in environment variables for the test window only. Anything you mark as secret is blanked before it reaches the model judge, and never printed in the report.
Which agents can you test?
Anything we can send a message to over HTTPS — a custom endpoint or an OpenAI-compatible API — whichever model it runs on. We test your agent: its prompts, retrieval and behaviour, not the underlying model.
How much of our time does it take?
About an hour in total: the kick-off call, answering a short checklist, and the readout. The testing itself runs on our side.
What does it cost?
Our first three audits are free, in exchange for honest feedback and — if you are happy with it — a short case study.
Find out what your agent does under attack.
Our first three audits are free. Tell us what your agent does and we’ll take it from there.