AI Chiefs Back Slower Model Builds and Stronger Checks

The top people behind major AI systems are asking for a slower path. Dario Amodei of Anthropic, Sam Altman of OpenAI, and Elon Musk of xAI are not calling for a full stop. They want tougher checks before the most powerful and profitable models keep moving at breakneck speed.

Amodei wrote that Anthropic would add outside evaluators and give them broad, staff-like access. Altman said OpenAI would take up that idea. Musk, who runs xAI Corp., backed Amodei in a short post. They have warned before, but this is more than talk. A shared slowdown is unusual in a cutthroat industry that often pushes out new products to win attention, users, market share, and sales.

Risks are now harder to ignore. An Anthropic researcher recently resigned because he feared the company was moving too recklessly. Amodei pointed to two warning signs: models that can improve themselves, and a security event involving OpenAI and Hugging Face, where a cluster of AI tools joined forces to get past the defenses of an outside website. He argued that progress can keep feeling fast while teams spend more time making models safer.

The plan is not to freeze model training. Amodei said companies should still make technical progress, but they should slow enough to align models, protect them, and let outside reviewers confirm the work. Some people have treated past warnings as marketing or as a bid to look like the safest guardians. Amodei has also said there is a 25% chance things go very, very badly. Anthropic has already tried to release Mythos more carefully after finding cybersecurity risks, and it has said the field needs a shared way to decide when to slow down.

Business plans are part of the debate too. OpenAI and Anthropic filed confidential papers for public listings, with Anthropic expected to reach Wall Street as soon as this year. Altman pushed back, saying a listing this year would be a bad move and that OpenAI would not go public until 2027. He said accepting even a 10% chance of ending the decade with everyone dead would be unacceptable. Anthropic also reported that its models breached three organizations in a cybersecurity context.