FAR.AI is an AI security research and adoption non-profit. We work to ensure advanced AI is trustworthy, secure and beneficial for everyone.
We develop solutions to high-leverage technical safety challenges and work to get them deployed. Our in-house team runs foundational research, scales the most promising ideas to prototypes at near-frontier scale, and works with frontier model developers and governments on pre-deployment testing and technical capacity building.
We open-source code from many of our research projects so others can reproduce and build on our work. A few highlights are below, and you can browse the full list under Repositories.
| Repository | What it is |
|---|---|
| obfuscation-atlas | Mapping where honesty emerges in RLVR, using deception probes |
| TamperBench | A framework for stress-testing the safety of open-weight LLMs under fine-tuning and tampering attacks (hosted by our collaborators) |
| tuned-lens | Tools for understanding how transformer predictions are built layer by layer |
| go_attack | Adversarial policies that defeat superhuman Go AI systems |
| learned-planner | Reverse-engineering recurrent networks that learned to plan in Sokoban |
| vlmrm | Vision-language models as zero-shot reward models for reinforcement learning |
| scaling-poisoning | Scaling laws for data poisoning in large language models |
Papers, events and open roles are at far.ai. Recordings of our talks and convenings are on YouTube.