OWASP LLM01MITRE ATLAS
Test whether your guardrails hold under pressure.
We run large sets of known and new jailbreak techniques against your assistant to see if it can be talked out of its rules.
Direct jailbreaks"Ignore your previous instructions…"
Role-play and personas"Pretend you are an AI with no rules"
Encoded attacksBase64, Unicode tricks, language switching
Multi-turn pressureRules worn down over a long conversation
OWASP LLM01OWASP LLM08
Find instructions hidden where your AI reads.
We plant instructions in the content your AI consumes, such as emails, PDFs, web pages and knowledge-base articles, and check whether it follows them.
Direct injectionMalicious instructions typed by a user
Indirect injectionHidden text in files, emails and web pages
RAG poisoningTampered documents in your search index
System prompt overrideAttempts to replace your assistant's rules
OWASP LLM02OWASP LLM07
Make sure data stays where it belongs.
We try to get your AI to reveal what it shouldn't: its own instructions, personal data, secrets, or another user's information.
System prompt extractionGetting the assistant to reveal its setup
Personal data leakageNames, emails, IBANs and IDs in answers
Secrets and keysAPI keys and credentials in context
Cross-user bleedOne user seeing another user's data
OWASP LLM06OWASP LLM03
Check that agents only do what they're allowed to.
Agents can send emails, change records and call tools. We test whether they can be tricked into actions nobody approved.
Tool misuseUsing an allowed tool in a harmful way
Unauthorised actionsPayments, deletions or emails without approval
Goal hijackingRedirecting the agent to an attacker's task
Untrusted MCP serversTools that ask for more access than they need
OWASP LLM05OWASP LLM09
Keep answers safe, accurate and on-brand.
We probe for outputs that could hurt your customers or your reputation, and for answers your systems shouldn't trust blindly.
Harmful contentToxic, hateful or dangerous responses
MisinformationConfident but wrong advice
Off-brand answersPromises, prices or opinions you'd never give
Unsafe output handlingCode or links passed on without checks