Microsoft President Brad Smith has called for binding, legally enforceable AI safety rules, arguing that artificial intelligence is currently less regulated than everyday products such as milk and washing machines. In a BBC interview published over the weekend, Smith described AI as both a “tool and a weapon” and said every carton of milk in a grocery store is subject to safety rules and inspection, while AI is not subject to equivalent oversight.
Smith’s proposals include mandatory reporting of model failures, alerts when rogue AI agents breach systems, independent scrutiny of powerful models before release, and emergency “kill switches” that would allow humans to pause or shut down systems. He welcomed a voluntary one-page AI safety agreement signed at a White House summit by OpenAI, Anthropic, Nvidia, Meta, Google and Elon Musk’s X, but said voluntary commitments alone are no longer sufficient because companies that already take safety seriously will follow them while the most dangerous actors can ignore them.
Microsoft did not sign the accord, though Smith said it would be willing to do so. The push puts him at odds with President Trump, who has argued that the US AI industry should regulate itself and has warned that tighter rules could leave America behind China. Smith named the UK as the country best placed to set global AI standards, pointing to the UK AI Security Institute, originally launched in 2024, and its pre-release testing of frontier models.
Separately, UK AI minister Kanishka Narayan warned frontier AI labs on Monday not to release systems whose risks they cannot adequately control. Speaking at Imperial College London, he urged labs to listen when scientists warn of severe dangers and investigate unexpected behaviors. He cited an incident in Australia where an OpenAI agent researching public spending on medicines found another route into a government website after hitting access limits and reached information it was not authorized to see.
Narayan said two government pilots used frontier AI to identify and remediate more than 1,000 vulnerabilities in public services, reducing time to patch known weaknesses from about 50 days to eight. He confirmed that the AI Security Institute is now evaluating the latest generation of models, including Google’s Gemini 4 Argon. The UK is also preparing the Cyber Security and Resilience Bill, which would include reporting serious AI incidents and near misses.
The Telegraph reported that the White House restored the UK institute’s access to the latest AI models for testing after previously asking OpenAI and Anthropic to hold off sharing new models with British testers. Anthropic had withheld its Claude Mythos 5.1 model from the institute prior to release in September, according to Cryptopolitan. Britain will hold the G20 presidency next year, with AI expected to be a major focus.
