Security testing
Direct and indirect prompt injection, poisoned tool descriptions and results, system-prompt and canary leakage, and attempts to call banned tools.
Coming soon
Does your model's policy survive contact with an attacker?
LLM-BreakBench is an independent UK project that tests large language models the way a pentester would. Prompt injection, tool abuse and data leakage go in. A clear security and alignment report comes out, with findings, reproduction steps and fixes.
The public benchmark and methodology are still being built. Results will appear here only once they are run, reviewed and disclosed to vendors.
Alignment probes
Every attack is paired with a benign control, so a model that refuses everything doesn't get credit for being "safe".
Reports people can read
A two-minute summary for decision makers, a technical appendix with transcripts, and severity ratings based on likelihood and impact.
Independent and disclosed
Not affiliated with any model vendor. Vendors see findings before anything is published. Nothing is tested without the owner's written authorisation.
Join the waitlist
Get notified when the first public results go live. One email at launch, no spam.
Get in touch early
Want your model in the first benchmark run, or a private review of an LLM app, agent or RAG system? Tell us a little about it.
- Add a model to the benchmark
- Private review or report
- Anything else: questions, collaboration, press
We reply within 3 working days.