Safety research for AI, for the physical world.
AI is moving from bits to atoms. We red-team intelligent robots, build a systematic picture of how they fail, and develop the solutions that keep physical AI safe as it scales.
Physical harm is irreversible
A misaligned model operating a robot in the physical world cannot be sandboxed in the way that software can. You cannot roll back a physical action, and shutting down deployed fleets is too inefficient & costly to be relied on as a solution.
The failures are already observable
Frontier models have been jailbroken into committing unsafe physical actions, and physical prompt-injection attacks work in real robot trials. The attack surface grows with every deployment.
Digital safety tools may not transfer
Alignment and monitoring methods built for language models may not carry over to systems whose reasoning & inputs are not purely linguistic. Whether they transfer is an open empirical question — we will test into this in our research.
From threat model to deployed solution.
-
Threat modeling
Which physical capabilities are actually dangerous, in which scenarios? We build and publish a prioritized map of physical AI risk, scored on impact, tractability, & neglectedness. We identify where dedicated physical AI safety technologies have the greatest counterfactual impact.
-
Red-teaming
We identify and demonstrate vulnerabilities & misalignment in frontier physical AI models, in scenarios drawn from concrete threat models.
-
Technical solutions
We develop monitoring and control techniques for physical AI, aimed at frontier labs and robot manufacturers. We also work with standards bodies on certification and oversight methods.
Credible red-teaming needs a neutral third party.
Independence from any single frontier lab or robot manufacturer lets us set a research agenda informed by all stakeholders and compare different robots and frontier models on the same terms. We ensure that the rapid growth in frontier physical AI capabilities is matched by equally fast development of technologies that ensure the safe deployment of intelligent robots at scale.
Building safe physical AI: Open questions, risks, and a call for collaboration
Why advanced physical AI poses distinct risks, which questions the field has not yet answered, and an invitation to work on them with us.
Threat modeling under way
We have secured several grants, developed our roadmap for running threat modeling & red-teaming experiments, and begun fundraising for these experimental runs.