Safety is an engineering discipline.
We started Reinforced Labs because the gap between how AI is tested and how it is used had become dangerous. Models were being given payment rails, inboxes and production databases, and evaluated with a spreadsheet of prompts someone wrote before lunch.
If it isn't measured, it isn't safe
Every safety claim we make, for us or for a customer, is backed by a test that can be rerun.
Safe and useful are one objective
A model that refuses everything has failed too. We optimise harm and helpfulness together.
Attackers adapt, so tests must too
Static benchmarks go stale. Continuous adversarial pressure is the only honest baseline.
Experts, not crowds
The hardest safety judgements need clinicians, scientists and security researchers. We pay for expertise and show our agreement numbers.
Researchers, red teamers and the people who ship.
Alumni of frontier AI labs and top university ML groups, with published work on adversarial robustness, reward modelling and evaluation methodology.
Former penetration testers, exploit developers and incident responders from national CERTs, a top-5 bank and major security vendors.
A network of 2,300 specialists in medicine, biosecurity, law, finance and 41 languages, who write and judge the hardest cases.
offices: San Francisco, London, Zürich
full-time researchers and engineers
frontier and enterprise model releases evaluated pre-launch this year
automated adversarial coverage for customers
Work on the problems that decide whether AI goes well.
We hire research scientists, offensive security engineers, platform engineers and domain specialists. If you have broken something important, or measured something nobody could, we'd like to hear from you.
Get in touch →Find out what your model does before your users do.
A 45-minute briefing with our evaluation team. Bring a use case; leave with a threat model, a draft eval plan and a sense of where you stand.