one year on
OpenAI and Anthropic sign pre-release safety testing agreements with US AI Safety Institute
The first-of-their-kind memorandums of understanding give the government access to major new models before public release, marking a formal turn from voluntary commitments to structured collaboration.
The US Artificial Intelligence Safety Institute (AISI) today signed first-of-their-kind agreements with OpenAI and Anthropic, giving the federal government pre-release access to major new models for safety testing and evaluation. The memorandums of understanding establish a framework for collaboration on research, testing and risk mitigation, the institute announced.
Under the deals, AISI will receive access to models both prior to and after public release, and will provide feedback on safety improvements in coordination with the UK AI Safety Institute. The agreements go deeper than the voluntary commitments the Biden administration has previously secured from 16 AI companies, formalizing government access and joint research.
Elizabeth Kelly, director of the US AI Safety Institute, said the milestone marks the start of technical collaborations to advance the science of AI safety. Anthropic’s Jack Clark noted the partnerships leverage the government’s expertise to ‘rigorously test our models before widespread deployment.’
The record
Director of the US AI Safety Institute said the agreements are 'just the start' but an 'important milestone' for responsible AI stewardship.
Anthropic co-founder said the collaboration leverages government expertise to test models before deployment, setting new benchmarks for safety.
One year later — open only if you can handle spoilers
In June 2025, AISI was re-established as CAISI by Secretary of Commerce Howard Lutnick.