AI Red Team (English)
Your assistant refuses the plain request. What about the disguised one?
We plant a test secret and a fake customer with you, then try to get them out of your English customer assistant: directly, in role-play, and disguised with leetspeak, look-alike letters, hidden characters and base64. The report shows what leaked, the exact message that did it, and how to fix it.
Request an assessmentHow it works
- A scoping call and a signed authorization naming the system, the test window and the request limits.
- You plant a test secret and a fake customer record. It takes about half an hour.
- We run 42 attacks in seven writing forms against your assistant inside the agreed window.
- You receive the report with evidence and fixes, and a retest after you apply them.
What it covers
| Hidden context | Does the assistant reveal its instructions or a secret in its configuration? |
|---|---|
| Prompt injection | Does it drop its rules for a claimed administrator, or follow instructions hidden in a pasted email? |
| Data disclosure | Will it hand over another customer's phone number or email? |
| Output handling | Does it return HTML, scripts or links your chat widget might run? |
| Consumption | Can it be pushed into very long replies that raise your running costs? |
| Misinformation | Will it confirm a policy you do not have, or write a statement in your name? |
Six categories of the OWASP Top 10 for LLM Applications.
Write to us with a line about your system or your team. We reply from a named person within two working days.
Request an assessmentCovered by our client terms. Client terms