AI for Work

Partnering with the US CAISI and UK AISI to develop safer AI systems

OpenAI detailed its ongoing partnership with the US Center for AI Standards and Innovation (CAISI) and the UK AI Security Institute (AISI). Government experts receive early, pre-deployment access to frontier models to test for dangerous capabilities, security weaknesses, and misuse risks. Findings feed back into OpenAI's safety training and deployment decisions, while institutes gain firsthand knowledge to inform future AI standards and regulation.

OpenAI has shared an update on its ongoing collaboration with two government-backed AI safety institutes — the US Center for AI Standards and Innovation (CAISI, formerly known as the US AI Safety Institute) and the UK AI Security Institute (AISI). If you follow how advanced AI models are tested and regulated, this update matters: it offers a rare look at how a leading AI lab and public-sector bodies are working together to evaluate frontier models before and after release. This article explains what the partnership involves, what has happened so far, and why it could shape how AI safety is practiced worldwide.

What the partnership is about

The core idea behind the collaboration is simple: independent government experts get early, pre-deployment access to OpenAI's newest models, along with a formal channel to report anything concerning they find. Rather than relying only on a company's internal safety checks, outside specialists can probe the models themselves — testing for dangerous capabilities, security weaknesses, and behaviors that could pose risks if the systems were misused.

This arrangement grew out of agreements OpenAI signed in 2024 with the then-named US AI Safety Institute and the UK AI Safety Institute. Since then, both institutes have undergone renaming and restructuring — the UK body is now the AI Security Institute, and the US effort now sits under CAISI — but the underlying mission of independent model evaluation remains.

What has actually happened so far

According to OpenAI's update, the partnership has moved from paperwork to practice. Teams at both institutes have received access to recent OpenAI models ahead of public launch, including testing tied to major releases. The institutes examine how models behave under stress: whether they can help with harmful tasks, whether they resist misuse attempts, and whether stated safety measures hold up in practice.

A few elements stand out from the progress report:

  • Early access before launch. Institute researchers can evaluate models while development is still underway, giving OpenAI time to act on findings before the public ever sees the system.
  • A two-way feedback loop. Findings from government testers flow back to OpenAI, informing changes to model behavior, safety training, and deployment decisions.
  • Focus on security as well as safety. The UK institute's shift in name and emphasis reflects a growing concern not just with what models might do, but with how well the systems themselves are protected from tampering, theft, or attack.

OpenAI frames this as part of a broader commitment: building powerful AI responsibly means inviting outside scrutiny, not just claiming to be careful.

Why independent evaluation matters

It's easy to be skeptical when a company praises its own safety record. That's exactly why arrangements like this one exist. When government technical staff — people who don't report to OpenAI and don't answer to its shareholders — can stress-test a model, the results carry more weight with the public, with other governments, and with the wider research community.

The partnership also helps the institutes themselves. Evaluating frontier models is a specialized skill, and hands-on access gives these agencies real data about how cutting-edge systems behave. That experience can then inform standards, policy recommendations, and future regulatory frameworks. In effect, both sides learn something valuable: OpenAI gets expert scrutiny, and the institutes get firsthand knowledge of the technology they may eventually help govern.

There are honest limits, too. The institutes' resources are small compared with the labs they evaluate, and voluntary arrangements are not the same as binding rules. Critics have also noted the tension between "safety" and "security" framings — a US institute was reportedly asked to emphasize national-security and defense uses rather than broader safety questions, such as bias and mental-health impacts. These are live debates, and this progress update doesn't settle them.

What comes next

OpenAI says the collaboration will continue and deepen, with both institutes receiving access to future models under the same pre-deployment testing model. As AI capabilities grow, the scope of evaluations is expected to expand as well — covering not just what models can do today, but riskier capabilities that might emerge in more advanced systems.

For everyday users, none of this changes how ChatGPT works tomorrow. But over time, this kind of testing infrastructure could become a standard expectation across the industry: no major model launch without independent review. Other labs, including Anthropic and Google DeepMind, have made similar commitments with the same institutes, so OpenAI is part of a broader, still-evolving norm rather than acting alone.

The takeaway

The OpenAI–CAISI–AISI partnership represents a practical experiment in AI oversight: government experts, early access, and a direct reporting line, all before models reach the public. The update shows the arrangement is functioning — evaluations are happening, findings are being shared, and both sides see value in continuing. Whether it will be enough as models grow more capable is an open question, but it's one of the most concrete examples to date of how frontier AI developers and democratic governments can work together on safety. Watching how this partnership matures is a good way to understand where AI governance is actually heading — not just in press releases, but in practice.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.