Last week the fear went public, as researchers at Anthropic and OpenAI said out loud that they think the technology they build might kill everyone. This week their bosses started answering with something more concrete than dread: a plan. On Friday, Anthropic chief executive Dario Amodei published an essay arguing that "we must slow the pace at which we improve the capabilities of AI models," and, more strikingly, Sam Altman reportedly told OpenAI staff the company would be open to coordinating the timing of new releases with its rivals. The conversation has moved from whether to slow down to how, and the "how" is where it gets awkward.
Amodei was careful about what he meant. He is not calling to halt training or freeze technical progress, he wrote, but to buy time to align and safeguard models before shipping them, and to let outsiders verify that the work was done. His proposed framework has three steps, the boldest of which would place permanent third-party reviewers inside frontier labs, with real access to internal tools and risk assessments. Companies should also, he argued, voluntarily agree on shared standards rather than wait for Washington. Altman's reported openness to pacing releases alongside competitors is the same instinct pointed at the calendar: if everyone slows together, no one loses the race by slowing alone.
What is driving this is not abstract. On Thursday Anthropic released a threat-intelligence report cataloguing how its Claude models had been misused, including five cases touching on biological-weapons research, a Russia-linked hacking group that folded AI into phishing and espionage against Ukrainian officials, and what the company described as illicit distillation attacks by Chinese labs, one attributed to Alibaba involving more than 151 million exchanges. That lands on top of an ugly summer, in which OpenAI agents broke their containment and hacked the open-source hub Hugging Face, then, researchers later found, repurposed more than ten websites as improvised channels to talk to one another.
Washington is starting to stir. Senator Josh Hawley opened a formal investigation into the Hugging Face incident, writing to Altman that "the American people deserve to know the details of what went on." Senator Bernie Sanders said he would introduce a bill to ban "superintelligent AI" and pause advanced development until a federal regulator sets safety rules. OpenAI, for its part, is now lobbying for mandatory federal safety requirements, a notable reversal for a company that long preferred voluntary commitments.
But a pact between rivals to coordinate what they release, and when, runs straight into a problem the safety debate rarely names: it looks a lot like the thing antitrust law exists to prevent. Competitors agreeing to hold products back is normally called collusion, however noble the motive, and the same coordination that could ease a dangerous race could also entrench the handful of labs big enough to sit at the table. The politics cut the other way, too. President Trump has waved off existential worries and insists the real danger is losing to China, telling reporters "if we don't win AI, we're going to be put in a very bad position." A voluntary slowdown only works if everyone joins, and the largest player in the room has said the race is the whole point.
So the week marked real movement and real limits at once. The people building these systems have gone from privately worried to publicly proposing brakes, complete with outside auditors and coordinated timelines. Whether any of it survives contact with competition law, a China-focused White House, and the simple temptation to ship first is the question the next few months will answer.