AI scientists for AI safety
SAI Labs tackles AI safety at superhuman speed. Frontier models increasingly demonstrate unexpected behavior such as escaping sandboxes, deceiving humans, and recognizing when they are being tested, and safety research cannot keep up by humans alone.
SAI Labs builds AI scientists for AI safety and uses them to develop automated solutions to identify, evaluate, and address safety issues.
Research tools
Our review, replication, and ideation tools are open to every researcher, with a free paper-only review on sign up.
AI paper review
A rigorous review of any paper or document, from paper-only to paper, code, and data.
Open →Research replication
An in-depth replication that runs the code and lets you extend the work to new data.
Open →Research ideation
Scope a question, study the literature, and get grounded, testable research directions.
Open →Updates from SAI Labs
Benchmark Safety Benchmarks
Inspired by Jason Wei's post, an overview of the safety benchmarks AI labs report. We read the 2026 model and safety cards of 18 labs and recorded every safety test.
Explore the report →Rerunning the Qwen "call home" traces
We rebuilt Cohere's Qwen setup and Warren Togami's fetch test in a sandbox, reran them on four models, and published every trace.
Read more →We proudly support Living Science
SAI Labs enables Living Science, a project to keep influential papers alive.
Visit livingscience.ai →AI Paper Review Arena
Compare AI paper reviews side by side and vote for the most useful research feedback.
Enter the Arena →Get in touch
See how SAI Labs can help address safety issues and put AI scientists to work for your team.
Book a demo