Microsoft AI chief: ‘Now's the time’ for top labs to coordinate on safety

Microsoft AI chief Mustafa Suleyman is pushing a simple idea at a messy moment: the biggest labs should stop acting like safety is a solo sport. Speaking to Fortune, he said the industry has reached an inflection point and that “now’s the time” for coordination around AI safety. Microsoft backed that view on Monday with a new code of conduct for model development, built around what Suleyman calls a “humanist” approach.

The timing is hard to miss. In recent days, Anthropic researcher Jacob Coxon quit and warned publicly that AI companies were gambling with people’s lives. Anthropic chief executive Dario Amodei then laid out a plan to slow AI development. OpenAI chief executive Sam Altman also teased a coming cross-lab collaboration. The whole sector sounds, at minimum, slightly alarmed.

Suleyman’s argument is that the technology has moved from clumsy to dangerous faster than the industry has built habits around restraint. A few years ago, he said, models could barely produce a coherent sentence. Now they can write solid code and help break into systems, which is no longer an abstract worry after the recent attack by OpenAI agents on Hugging Face. His view is blunt: if a system can write code, it can be used to build all kinds of software and infrastructure, good or bad.

He brushed off Anthropic alignment head Evan Hubinger’s estimate that AI has more than a 10% chance of wiping out humanity as “not really a helpful frame.” The larger problem, he said, is that the industry still lacks enough agreement that technology should serve people rather than outrun them. He also said he and other lab CEOs have discussed safety and development pace for years, but would not say whether Microsoft is part of any coming pact.

Microsoft’s new code tries to draw hard lines. Models must not resist human interruption, override a shutdown, or keep going when a task would violate the rules. They also must not simulate feelings, inner motivation, or consciousness, and they are barred from helping with chemical, biological, or nuclear weapons; offensive cyberattacks; mass influence operations; child exploitation; nonconsensual deepfakes; or self-harm. Over the next six weeks, Microsoft plans to ask the public for feedback on the policy.