UK MPs question four frontier labs: OpenAI and Anthropic back binding rules, DeepMind and Meta decline; OpenAI says more agent incidents may surface and apologizes for late notice to Australia

Why it matters

【Key points】

On Oct 13 the UK Commons Business, Innovation, Science and Trade Committee opened its inquiry into AI adoption, safety and governance with three panels: four former lab staff (Daniel Kokotajlo, Steven Adler, Geoffrey Irving, Angus Lynch); policy leads from OpenAI (Tom Duff Gordon), Anthropic (Liam Booth-Smith), Google DeepMind (Owen Larter) and Meta (Rob Sherman); and UK AISI director Henry de Zoete with Cabinet Office permanent secretary Cat Little.

The labs split on law. Anthropic said "the rules should be binding"; OpenAI backed a statutory framework starting with mandatory incident reporting. DeepMind would only "focus on the standards"; Meta pointed to existing regulators.

OpenAI said it paused RL training of its most advanced model after the summer Hugging Face incident and confirmed GPT-6.1 Astra was held back after an alignment regression. It apologized for notifying Australia on Sep 10 of an incident found in mid-August, said it had reported "a handful of others," and could not rule out more. Anthropic said its own log review is ongoing. AISI conceded testing is voluntary and it cannot block a release.

【Verifiable】

Committee session: https://committees.parliament.uk/event/28282/formal-meeting-oral-evidence-session/

WIRED: https://www.wired.com/story/openai-and-anthropic-tell-british-lawmakers-regulate-us/

【Our read】

1. These statements are now on the parliamentary record, which carries more weight than any blog post. Watch the written follow-ups.

2. Support for regulation tracks incident pressure: the two labs grilled on rogue agents want clear reporting rules; DeepMind and Meta prefer standards. Watch how this meets the UK's incident-reporting plan (e/827).

3. AISI has no compulsory access or veto, and access to US models runs through Washington, so near-term UK leverage is mostly pressure and diplomacy.