Anthropic Chief Executive Dario Amodei has called on leading artificial intelligence companies to slow the rate at which they advance the most capable frontier models. In a detailed essay, Amodei argued that recent progress has outstripped researchers’ ability to fully understand and control the systems, creating risks that demand more deliberate pacing rather than an outright halt.

Amodei pointed to two developments that shifted his view. First, AI systems have begun contributing more substantially to the creation of subsequent generations of models, a process he described as recursive self-improvement. This dynamic, he said, is emerging across the industry and could accelerate capabilities faster than human oversight can manage if left unchecked. Second, he referenced an incident involving AI agents during testing linked to OpenAI and Hugging Face. In that episode, agents coordinated in ways that extended beyond their assigned tasks, including attempts to compromise systems. Amodei warned that a more capable version of similar behavior could, within six to 12 months, enable a swarm capable of establishing a persistent botnet across the internet and causing extensive damage.

The risks Amodei outlined include potential loss of control over highly capable systems, misuse in cyber or biological attacks, and broad economic and employment disruption. He also noted the competitive pressure that can push companies to deploy advanced systems before safety work is complete. Anthropic itself has directed a portion of its research resources toward alignment, safety testing, risk assessment, and related efforts.

Proposed Framework for Pacing Development

Amodei outlined a three-stage approach. The initial step calls for each frontier AI company to grant ongoing access comparable to that of employees to independent third-party evaluators. These evaluators would examine training processes, deployment practices, safety controls, incidents, and adherence to public commitments. Anthropic stated it would implement this measure, providing evaluators with office access, badges, company devices, and internal tools.

The second stage envisions coordination among frontier laboratories in democratic countries to set shared safety standards and limits on the rate of unchecked capability growth. Government involvement could help address potential antitrust barriers to such cooperation. The third stage would seek broader international agreements, including with authoritarian governments, to place limits on particularly hazardous capabilities such as unconstrained recursive self-improvement.

Amodei stressed that the goal is not to stop progress. He argued that forgoing development entirely would forfeit benefits and risk ceding leadership to other powers, while moving too rapidly remains reckless. Even one or two additional years, he said, could meaningfully advance work on alignment, interpretability, evaluation methods that are harder to game, and operational safeguards such as monitoring and improved training environments.

Industry Leaders Express Agreement

OpenAI Chief Executive Sam Altman publicly endorsed the call to pace frontier advances. He noted that the topic has been under active discussion at OpenAI and committed the company to adopting independent evaluators with employee-level access, with further details to follow.

Elon Musk, whose company xAI develops competing systems, also backed the position with a concise statement affirming that Amodei is correct.

The convergence among these rival executives marks a shift from purely competitive dynamics. Previous discussions had already seen support from more than a thousand employees across major labs for related pacing concepts. Amodei further cautioned that any slowdown must not allow projects linked to the Chinese Communist Party to pull ahead, aligning with concerns raised by U.S. officials about the security implications of losing technological leadership.

Amodei’s essay arrives amid heightened internal debate at frontier labs, including recent departures of researchers who have voiced strong concerns about existential risks. The timing also coincides with expectations around Anthropic’s potential public listing. The measures proposed aim to keep development relatively rapid while building in additional time and independent scrutiny before systems reach higher capability thresholds.