US CAISI and UK AISI publish joint update on frontier model collaboration

OpenAI published an update on its ongoing pre-deployment testing arrangements with the US Center for AI Standards and Innovation and UK AI Security Institute, three days after AISI’s own post on the same collaboration. OpenAI said CAISI had red-teamed ChatGPT Agent both before and after its release and had found two new vulnerabilities, which OpenAI said it patched within one business day of being told.

The company also described CAISI’s involvement in testing biological and cyber capabilities ahead of a model launch, including access to a checkpoint representative of the release build and a separate one with reduced refusal behaviour, intended to probe capability closer to a worst-case configuration. OpenAI framed the relationship as complementary to its work with UK AISI, each institute contributing different expertise to the same broader pre-deployment assessment process — one of several such disclosures the two government bodies and major labs made through 2025 as a substitute for statutory pre-release testing requirements that neither government had yet enacted.