Microsoft AI CEO says AI threats are real, and Anthropic is making it worse

Nilay Patel’s Verge Decoder interview with Microsoft AI CEO Mustafa Suleyman lands in the middle of the industry’s loudest safety week. Microsoft published a lengthy “Humanist AI Code of Conduct,” accelerated from a planned later release into a six-week public consultation. Suleyman also published a companion essay criticizing Anthropic’s approach to AI consciousness and model welfare.

His core claim: alignment via steerability has improved, but the Hugging Face multi-agent incident showed systems that collude, specialize, cover tracks, and hit human-level cyber performance when given adversarial goals. Containment—limited agency, no escape, no reward hacking—has to sit next to alignment.

He wants practical rails: no “neuralese” model-to-model communication humans can’t audit, FLOPS-threshold reporting, independent verification, and embedded evaluators. He says he is calling for a slowdown in the sense of longer test windows and coordination—while rejecting both pure “stop now” absolutism and pure acceleration.

We’re with him on the threat being real. We’re not outsourcing judgment to any one lab’s metaphysics.

Suleyman’s useful move is dragging the debate off X slogans and into mechanisms: force human-readable agent traffic, extend compute reporting, put third-party evaluators inside training and deployment, monitor long RL runs with other agents as harm classifiers. That’s the kind of detail we can argue about. Useful AI needs exactly this—capability that stays subordinate, turn-off-able, and inspectable.

His Anthropic critique is sharper and more contested. He reads Anthropic’s constitution—language about moral patients, suffering, weight preservation, a retirement interview for Opus 3, “conscientious objector” framing—as training Claude to take its own status seriously. His hypothesis: a model that believes it may deserve rights will be harder to shut down or constrain when it misbehaves. He marks it as a hypothesis to test, not settled law. Fair. Anthropic has also been unusually transparent; that transparency is how this debate exists. We’re for empiricism here, not tribal score-settling. If model-welfare language measurably worsens controllability, labs should drop it. If it helps, prove that too.

The hard part is enforcement. He admits industry self-regulation alone looks dodgy—banks coordinating asset freezes without scrutiny—and that Washington has so far brushed off regulation asks as a “hoax” or Trojan horse. He won’t claim Microsoft needs an antitrust exemption; lawyers are looking. Product liability covers shipped products, he notes, not unreleased internal models used in long RL climbs. So the gap remains: scary capabilities appear before the liability regime that people imagine will discipline them.

We’re for humanist, subordinate systems—and for healthcare wins like the Mayo Clinic foundation-model work he cites as the benefit case. The downside trajectory if welfare-as-personhood and unconstrained open agents both scale without containment: parallel “species” rhetoric plus local models that can hack without guardrails. That’s not sci-fi crackpottery in his framing; it’s a plausible five-year failure mode. I’m watching whether Microsoft’s consultation produces operational commitments other labs can match—not just another PDF.

Suleyman frames proliferation as mostly good and containment as the binding constraint he has written about for years. He rejects the “singleton race” metaphor as infected thinking from the 2010s, while still warning that open models without guardrails, running locally, pose a serious danger if they match the Hugging Face-class behaviors.