Safety research has to move at the speed of capabilities.

AI capabilities are advancing faster than the tools we use to make sure those systems are safe. The work of evaluating models, finding failure modes, and building defenses is still largely manual — and the people doing it are outnumbered.

Neolithic builds open-source, agentic tools that give safety researchers leverage: tooling that takes on the repetitive, brittle parts of the work so researchers can focus on the hard questions — and that keeps results trustworthy as models grow more capable.

The long-term goal is to automate safety research and defensive systems end-to-end, so that safety can keep pace with — and eventually outpace — AI capabilities. We build in the open and as a nonprofit on purpose: this is infrastructure the whole field should be able to adopt and trust.