Brad Smith calls for global AI safety rules: independent evaluations, mandatory incident reports and an emergency brake

Microsoft President Brad Smith says the "moment has arrived" for global safety rules for frontier AI models. In a BBC interview released Monday, October 12, 2026, and in remarks to reporters in Dublin on October 13, he called for outside evaluation of models before release, mandatory incident reporting, a legal duty to tell people when an AI agent has harmed or breached them, and "safety brakes" that can stop a system. "Right now, your washing machine is subject to more safety rules than AI," he said, according to PYMNTS and The Eastern Herald.

This story is based on reporting by TechCrunch, AFP, Cybernews, Quartz and PYMNTS of Smith's BBC interview and Dublin remarks, checked October 14, 2026. Microsoft hasn't published a written version of the proposals on its On the Issues blog, where Smith's most recent post is dated September 23.

The short version

• Who and where: Brad Smith, Microsoft's vice chair and president, on a BBC podcast released October 12, then speaking to reporters on October 13, ahead of Dublin's International AI Summit on October 14.
• What he wants: independent evaluation before release, mandatory incident reports, notification of anyone harmed by an AI agent, and an emergency brake.
• Who should lead: the UK, which he says Washington views differently from Brussels.
• What it isn't: a Microsoft product commitment or a formal policy paper. These are proposals made in interviews.

What exactly did Brad Smith propose?

Pulling together the BBC interview (as reported by TechCrunch, Quartz and PYMNTS) and the Dublin remarks (as reported by AFP and Cybernews), Smith's ideas fall into four groups:

1. Independent evaluation before release. Outside evaluators would vet frontier models before launch. In Dublin, Smith questioned how that's being done today. He described a "mood in Silicon Valley" in which AI companies choose and pay their own evaluators, and said: "That does not sound like a recipe for independent and effective evaluation," according to Cybernews. He added: "There needs to be transparency for the findings from these evaluations."
2. Mandatory incident reporting. AI providers would have formal obligations to report incidents, AFP reported.
3. Notifying the people affected. Organizations whose systems are breached by rogue AI agents would have to be told, TechCrunch reported. AFP said Smith compared this to data-breach rules that require businesses to notify consumers: "The same logic should apply to people harmed by an AI agent."
4. An emergency brake. AI agents should have "safety brakes" that users can trigger after an incident, AFP reported. Smith said systems running critical infrastructure should be able to be turned off or slowed down. "We should put a safety break as close as we can to the user of the system," he said, according to Cybernews.

His broader argument was about consistency with other products. "We know how to do this for other products," he said, as quoted by The Eastern Herald. In Dublin he said, "We need to act with urgency," and called for "real and effective safety standards" the public can have confidence in (AFP).

Why does he think the UK should lead?

Smith told the BBC that the UK, not the US or EU, is best placed to drive global rules. "In Washington, D.C., people don't look at the U.K. and they don't look at London the way they look at Brussels," he said, according to TechCrunch. "That's a strength."

He pointed to the UK's AI Security Institute, which gets pre-release access to major US models. That access had been in question. Politico reported on September 24 that the White House asked OpenAI and Anthropic to hold new models back from British testers until a US review. On October 12, UK AI minister Kanishka Narayan said the institute was still running "rigorous evaluations of the latest generation of models" (our UK story).

What did he say about voluntary commitments?

In late September, leading AI companies signed a White House accord that President Trump called "morally binding" and that, according to TechCrunch, commits them to work with "independent auditors" on model safety. It has no formal enforcement. Smith called it a "good step" but said it falls short, because voluntary pledges mainly bind "good companies." "It's the companies that are not as responsible that you need to worry about," he said, according to TechCrunch.

His comment on evaluators came the same day Reuters reported that Anthropic and OpenAI had sought significant control over what outside evaluators could publish, as recently as October 5. Cybernews linked the two. That dispute is covered in our story on labs and evaluator independence.

How does this compare with Satya Nadella's 'emergency brake'?

Smith's emergency brake echoes his boss. On October 10, Microsoft CEO Satya Nadella wrote on X that companies should "assume a model is compromised" and that "an authorized person should always be able to pause or shut down a model mid-task" (our Nadella story).

The difference is the audience. Nadella set out principles for how companies should deploy models. Smith is asking governments to write rules, applied internationally, with mandatory reporting and notification duties. Neither man has said how an off switch would work technically for a model that's already deployed across many customers.

What else is happening?

The proposals overlap with others made the same week:

• On October 13, about 400 scientists including Geoffrey Hinton and Yoshua Bengio called for an international body to set containment standards and track containment failures (our story on the open letter).
• EU tech chief Henna Virkkunen told Reuters on October 9 that the EU AI Act is sufficient to address risks from rogue AI systems.

Smith isn't neutral. Microsoft sells AI models and agents, and governance tooling such as Microsoft Agent 365. Rules that require evaluation, logging and incident reporting favor companies with the resources to comply.

What's confirmed, reported and uncertain

• Reported (multiple outlets): the BBC interview's October 12 release and the "moment has arrived" and washing machine remarks; the UK leadership argument; the four proposals; the Dublin remarks on October 13.
• Single-outlet quotes: the "recipe for independent and effective evaluation" and "safety break" quotes come from Cybernews; the "people harmed by an AI agent" line from AFP.
• Uncertain: whether Microsoft will publish a formal proposal, who the independent evaluators would be, what powers they'd have and who would pay, and whether any government will take up the ideas. Smith was scheduled to give a keynote titled "Growth and Opportunity in the Intelligence Age" at Dublin's International AI Summit on October 14, according to the summit's media guide. As of our check on October 14, no report of what he said on stage had been published, and Microsoft hadn't posted a text of it.

What it means for you

Nothing changes legally today. But mandatory incident reporting and notification duties, if they come, would reach companies that deploy agents, not just the labs that train models. Teams running agents should already know what their agents can reach and keep logs that would support an incident report. Our case against unattended agents covers the controls that make that possible, and our prompt injection explainer covers one common way an agent is turned against its owner.