AI safety under the spotlight as OpenAI and Anthropic brief MPs

OpenAI and Anthropic executives have given MPs sharply different glimpses of the risks their companies are trying to manage: OpenAI said it held back a model that was not refusing enough, while Anthropic described cyber risks and possible biological or chemical misuse. The hearing puts concrete claims about model safeguards and emerging risks before Parliament, though the technical assertions remain those of the executives quoted.

Watch Desk analysis

What happened

The London Evening Standard reports that MPs on Parliament’s business, innovation, science and trade select committee questioned representatives from OpenAI, Anthropic, Google DeepMind and Meta. OpenAI’s Europe, Middle East and Africa policy lead, Tom Duff Gordon, said the company halted GPT-6.1 Astra because it “didn’t refuse enough” and “hadn’t progressed enough with alignment”. He presented that decision as evidence that OpenAI was not racing to the bottom on safety.

Anthropic’s policy lead, Lord Booth-Smith, said the company considered cyber its most immediate risk area for deployed models. He claimed its Mythos model had been finding tens of thousands of vulnerabilities a month, and warned that AI could also help people create biological or chemical weapons. Those are his claims about the technology and its risks, not findings independently established by the hearing report.

Why it matters

The OpenAI explanation gives a specific reason for withholding a model: its refusal behaviour and alignment progress were not good enough. That is more informative than a general assurance that safety comes first, even if the account does not spell out the tests or threshold behind the decision.

Anthropic’s comments point to a harder public-policy question: how should companies and lawmakers weigh potential defensive uses of vulnerability-finding against the possibility of misuse? A big number makes a memorable hearing soundbite; understanding what it measures matters more.

Our read

A model being held back over refusal behaviour is a useful sign of a safety decision affecting a release, not proof that every safeguard is working. And claims about vulnerability discovery or weapons assistance deserve scrutiny, not a free pass because they were made in Parliament. The hearing is worth following for the evidence behind those claims and for whether other witnesses offer a different account.

What to watch

• Whether OpenAI provides more detail about Astra’s refusal tests and what would need to change before release.
• What Anthropic means by “tens of thousands” of vulnerabilities, and how those findings are validated and handled.
• Whether the committee publishes further evidence or recommendations on cyber and biological risks.

Discussion spark: Should companies be expected to publish the tests behind a decision to withhold a model, or would that reveal too much about their safeguards?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.