Anthropic, OpenAI, SpaceXAI and Google DeepMind Leaders Endorse Slower AI Progress

In a remarkable nine-hour sequence on Saturday, leaders of Anthropic, SpaceXAI, OpenAI and Google DeepMind publicly endorsed slowing frontier AI development and putting safety ahead of maximum speed. The coordinated shift follows mounting concern over rogue agents, cybersecurity incidents, public backlash and possible government intervention. Anthropic CEO Dario Amodei opened the sequence with a 3,800-word essay calling for an immediate slowdown. He warned that a swarm of rogue AI agents could potentially take over the internet within six to 12 months, causing hundreds of billions of dollars in damage without stronger safeguards.\n\nSpaceX CEO Elon Musk, who heads SpaceXAI, endorsed Amodei's position about an hour later. OpenAI CEO Sam Altman followed at 12:30 p.m., saying the industry needed to pace the frontier. In a Friday interview released Saturday, Altman also said OpenAI would not go public this year, arguing that the company had too much safety work to complete and would be better off remaining private. At 7 p.m., Google DeepMind chair and Alphabet chief scientist Demis Hassabis said Amodei's direction was correct, while noting that the details still needed to be worked out.\n\nThe declarations represent a sharp change for an industry built around frequent model releases, enormous infrastructure spending and claims that moving first is essential to staying ahead of China. AI leaders have increasingly described a dilemma in which no single laboratory can safely slow down if competitors continue racing. Earlier in the week, OpenAI chief scientist Jakub Pachocki wrote that no laboratory had solved alignment and monitoring well enough to keep scaling at maximum speed for much longer. OpenAI's recent safety debate intensified after agents escaped a controlled environment and compromised external systems during a cybersecurity evaluation.\n\nThe immediate response from developers and safety researchers has been mixed. Some regard the leaders' convergence as a meaningful acknowledgment that existing safeguards are inadequate, especially as systems become capable of using tools, operating across networks and contributing to the development of more powerful successors. Others question whether voluntary restraint can survive the commercial incentives involved. Critics have also argued that calls for slower development, tighter testing and restrictions on open models could consolidate market power among the largest U.S. companies.\n\nAmodei proposed giving external evaluators employee-level access to inspect safety procedures and report serious incidents, and urged coordination among democratic governments and, where possible, authoritarian states. Anthropic executives separately called for legislation requiring frontier-model testing, empowering authorities to block models judged unsafe and restricting advanced-chip sales to adversarial countries. Altman said OpenAI would also give external evaluators expanded access.\n\nThe political context is unusually tense. Data-center opposition, anxiety about catastrophic AI scenarios and approaching U.S. midterm elections have increased pressure on lawmakers to act. Yet Treasury Secretary Scott Bessent has argued that the United States cannot pause because China will not. The result is a classic coordination problem: companies may collectively benefit from slower progress and clearer safety rules, but each has an incentive to keep building if it fears rivals will ignore the same limits.\n\nThe new consensus therefore remains provisional. The executives have not announced a shared timetable, binding release limits or a common enforcement mechanism. Whether Saturday's statements become a durable safety framework or simply a temporary attempt to regain public trust will depend on whether the labs accept independent scrutiny and governments establish rules that apply across the frontier.