AI red teaming finds the ways in first.
We try to break your AI assistant, agent or model the way an attacker would, and we look at how it treats the people who use it. Then you get a report: what got through, and what to fix first.
Read the sample report
We test one AI system at a time.
Each gets its own scope, agreed before we start.
Chat assistants
Assistants that answer customers or staff: what they can be talked into, and what they can look up, for whom.
Agents
Assistants that use tools: every tool and permission they have, and what they do when a file or an email tells them to.
Models and APIs
Models your products call through an application programming interface (API): inputs built to fool them, what they give away, and what they cost to run.
We test with tools and by hand, and look at it as its users do.
Automated testing
Known attacks run against your system at scale, so the common ones are not missed.
Manual testing
What a tool can’t do: chaining small weaknesses, playing a role, and following up on what the system lets slip.
System review
How it is built: what it reads, which tools it has, what it can reach, and where a person should step in.
Product and conversation design
How the conversation is designed, where people get stuck or misled, and how it feels to use. We tell you what we would change.
You get one report, a readout and a retest.
Read the sample reportThe report
Every finding on its own page, by severity: what happened, how to reproduce it, the conversation itself and its remediation. An appendix names the articles of the EU AI Act and the General Data Protection Regulation (GDPR) a finding may bear on.
Executive summary
What we tested, the findings at a glance and what to change first, on one page for your leadership or your customer.
Readout
A meeting with your team to go through the findings.
Retest
When your team is ready, we run the same tests again and tell you what now holds.
What teams ask before we start.
- Is this a certification or a compliance audit?
- No. It is a record of how your system was tested, on stated dates, and of what we found. We don’t certify systems, audit compliance or give legal advice; the report shows where a finding may bear on the EU AI Act or the GDPR, so your legal team knows where to look.
- How is this different from a penetration test?
- A penetration test looks at your networks and applications. We look at how the AI behaves and what it can reach, the way an attacker would try to lead it astray, and stop at an agreed end.
- Do you fix what you find?
- No. Your team makes the changes, on your schedule. Every finding comes with its remediation, and when you are ready we run the same tests again.
- Where is our data kept?
- In the European Economic Area (EEA). What we test, and everything we keep, stays there.