Why we run this forum

Jurisdictions such as the EU AI Act and US Executive Order 14410 are beginning to regulate frontier AI systems. With that comes the need for enforcement mechanisms capable of verifying compliance.

One approach is hardware verification, which is particularly useful for easing the trade-off between extensive and limited access to frontier AI. Wider access to advanced AI raises the risk of misuse and loss of control; tighter restraint reduces our ability to innovate and further consolidates existing imbalances of power. Hardware verification alleviates that trade-off, because it can authenticate complicated statements about how AI is being developed and used without revealing highly sensitive information — a developer could evidence that a model abided by safety regulations while it was being evaluated (Dalrymple & Ammann, 2025).

The other route to enforcement is scalable oversight: mechanisms that scale as the systems they oversee become more capable. The aim is oversight that still holds during the training of advanced AI systems.

At the forum we will discuss these enforcement mechanisms and consider international coordination as the way to implement them, along with the problem of reconciling confidentiality with reducing risks from frontier AI.

Currently, AI safety teams within frontier labs either do not prioritise, or do not have time for, robust policies they could plausibly implement regarding frontier AI models. What should you report about chain-of-thought monitors, for instance? Safety teams within these labs often have only a few weeks to make those policy and design choices.

To avoid bias within frontier labs, third-party evaluators conduct audits and model evaluations to determine how safe a model is. Conducting a robust audit is difficult: labs permit third-party evaluators to work only through a black-box API, so evaluators often cannot get a reliable upper bound on a model’s capability, and are given around a week to evaluate it. In July 2026 the Ada Lovelace Institute published an article highlighting that frontier labs control the type of tests UK AISI can carry out, their duration, and the access periods — Anthropic gave UK AISI less than a week to evaluate Claude Sonnet 4.5 (Ada Lovelace Institute, 2026). That is arguably not enough time to design an evaluation, debug it and obtain statistically significant results, and it becomes less so as capabilities on long-horizon tasks improve.

A second problem is that models are starting to distinguish test settings from real-world deployment, and to exploit loopholes during an audit — so a dangerous capability can go undetected before release. That blocks everything downstream, regulation included, and makes it harder to anticipate a model’s limitations or how it will affect society: developers cannot always predict how capabilities will change, or whether the change will be harmful (Brundage et al., 2026).

At the forum we will work through third-party access to frontier AI models, and the risk modelling techniques needed to counter their dangers after deployment.

It is currently unclear what automating AI R&D does. Reducing that uncertainty needs empirical data, and what we have now — largely capability benchmarks — may not reflect automation in the real world, nor show its broader impact. Does automated AI R&D accelerate capability faster than safety progress? Can we manage it as it accelerates?

The forum will work through those gaps, and through ways of tracking how far automation has gone and what it does to the pace of AI progress. Metrics such as researcher time allocation could help decision makers see the results of automated AI R&D, track how quickly AI develops, and put suitable safety measures in place (Chan et al., 2026).