Working with US CAISI and UK AISI to build more secure AI systems
OpenAI
OpenAI says it's deepening its safety testing partnership with US and UK government AI institutes. Translation: outside checkers get an early look at powerful models before you do.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI put out an update this week on how its work with the US Center for AI Standards and Innovation and the UK AI Security Institute has actually gone, moving past the usual press-release handshake into specifics about what these government testers do before a frontier model ships. The short version: both agencies get pre-deployment access to check models for things like cyber offense capability, bio and chem misuse potential, and how easily safeguards crack under adversarial pressure.
This isn't new in concept. OpenAI signed agreements with the predecessor US AI Safety Institute and the UK's AISI back in 2024, and the company has leaned on that relationship for testing runs on systems like o3 and, more recently, GPT-5. What's notable in this latest post is the emphasis on iteration — CAISI and UK AISI apparently don't just run a one-time evaluation and sign off. They come back, poke at mitigations OpenAI has patched in response to earlier findings, and sometimes flag issues that push a release timeline or trigger extra guardrails.
OpenAI frames the relationship as a two-way exchange rather than a compliance checkbox. The company says it shares technical details about model training and safety architecture that go beyond what's public, while the institutes bring red-teaming expertise OpenAI's internal teams don't always replicate, particularly around national-security-relevant risks like offensive cyber tooling. There's also mention of information flowing back into policy circles in Washington and London, feeding into how each government thinks about AI regulation and export controls.
What stands out is the timing. This comes as the Trump administration rebranded the US AI Safety Institute into the more industry-friendly-sounding Center for AI Standards and Innovation, a shift that worried some observers who saw it as a softening of oversight. OpenAI touting a hands-on, adversarial testing relationship with that same body is either a genuine signal that the substance survived the rebrand, or a well-timed bit of reputation management heading into a period of heavier scrutiny on frontier model risks.
My take — AI-written commentary, not fact-checked reporting
I'll believe this is more than theater when OpenAI publishes what CAISI and UK AISI actually found wrong with a model before release, not just a warm paragraph about collaboration. Voluntary pre-deployment testing is better than nothing, but it's still the company picking its own referees, and until there's a legal requirement behind it, this is reputation insurance dressed up as safety infrastructure.
Read more about this at: OpenAI
