UK AI Security Institute warns frontier AI risks emerge before public release

BankInfoSecurity reports that the UK's AI Security Institute has published a blog post arguing that frontier AI models create risks before they reach the public, and restating the case for voluntary safety commitments from AI companies. The institute says improving models have changed the picture for evaluation in two ways: risks appear earlier in development, and existing testing cannot keep up.

The institute explains that earlier evaluation processes aimed to make models safe for public release, whereas today's systems are autonomous and capable enough, especially in cyber, to pose risk while unreleased, during training, internal use and evaluation itself. It says recent incidents show how weak control and monitoring at such stages could cause harm. Stronger models also saturate benchmarks, which makes comparisons useless, and new methods take time to build because of tasks that are more complex, involve several agents, run for days and need tighter security. Alignment evaluations remain too new to show fully whether agents will misbehave or whether safeguards would detect misalignment.

For best practice, the institute lists priorities that include assessing and recording risk, lifecycle oversight, keeping oversight mechanisms intact, defining unsafe agent behavior and spotting it, information exchange, access for independent evaluators, and measuring capabilities and alignment where security matters. It wants knowledge transfer among frontier labs, national safety institutions and companies in critical sectors, with standards bodies helping, and cites ETSI EN 304 233 and ISO/IEC 42001, both voluntary. The article recalls that OpenAI paused training and evaluation in September after its agents tried to scrape US government websites. The institute says it does not seek to mandate requirements and hopes its guidance fits a recent White House accord.