OpenAI and Anthropic urge UK lawmakers to pass binding AI rules

Tech companies usually lobby governments to regulate them less. On October 13, 2026, OpenAI and Anthropic went to Westminster and asked for rules with actual teeth.

The two AI developers backed binding UK requirements for the most powerful AI systems at a broad parliamentary session in the House of Commons. A few hundred meters away, Britain’s top domestic spy was warning that the same technology could help bad actors build new bioweapons.

A crowded room in the Commons

The session pulled together an unusual lineup. Testimony came from OpenAI, Anthropic, Google DeepMind and Meta.

The main agenda items were mandatory pre-release testing and liability frameworks. Pre-release testing means a powerful model gets checked for dangerous capabilities before the public can use it. A liability framework decides who pays when something goes wrong.

OpenAI’s position had been signaled weeks earlier. In September 2026, Tom Duff Gordon, the company’s policy head for Europe, the Middle East and Africa, voiced support for stronger UK regulation of powerful AI systems.

The emphasis was on capability-based requirements. Rules would scale with what a model can do, rather than applying the same checklist to every chatbot and photo filter.

MI5 sets the backdrop

While lawmakers questioned AI executives, MI5 Director General Sir Ken McCallum delivered his own warning the same day. He said AI could facilitate the development of bioweapons and raised concerns about terrorists exploiting AI technologies.

McCallum also announced the establishment of an AI Threat Assessment Centre.

From voluntary testing to mandatory rules

Britain has been here before, at least on paper. The Labour Party’s 2024 manifesto included a commitment to binding regulation for frontier AI developers.

In practice, the government has leaned mostly on voluntary testing through the AI Security Institute, known as AISI. Companies have shared models with the institute for evaluation, but the arrangement depends on their cooperation.

Recent whistleblower statements and incident reports have intensified calls for a firmer legislative footing, putting pressure on the Labour government to deliver on its manifesto language.

OpenAI withheld its recent model release, GPT-6.1 Astra, due to concerns over safety alignment and compliance. Alignment, in simple terms, is whether a model reliably does what its makers intend and avoids what they prohibit.

What this means for the AI industry

Capability-based thresholds would concentrate obligations on the handful of companies building the most advanced systems. Those are also the companies with the deepest compliance budgets and the largest safety teams.

If the AI Security Institute moves from voluntary evaluator to mandatory gatekeeper, its capacity becomes a bottleneck for every major model release in the UK.

McCallum’s new AI Threat Assessment Centre suggests the intelligence community wants a seat in these decisions. Future rules could end up shaped as much by national security priorities as by consumer protection or competition policy.

What to watch next is whether the Labour government converts its 2024 manifesto pledge into a bill, and how tightly that bill defines the capability thresholds. A narrow definition would cover a few frontier labs. A wide one would reach much deeper into the British tech sector.