Anthropic told the White House it was holding back its most powerful AI model after safety tests produced results the company could not confidently clear, according to reporting by The Associated Press. White House Chief of Staff Susie Wiles met directly with Anthropic CEO Dario Amodei in spring 2025 to discuss the model’s risk profile, a conversation that elevated what might have been a quiet product delay into a matter of national concern.
The restricted release marks the first time a major AI company has publicly pointed to external biological risk research as a reason to limit access to a frontier model, and the first time such a decision has prompted a documented meeting at the highest level of the executive branch.
The biological risk study at the center of the decision
A key piece of evidence behind Anthropic’s caution is a research paper titled “Measuring Mid-2025 LLM-Assistance on Novice Performance in Biology,” published on arXiv. The study tested whether people with no formal biology training could use large language models to carry out technical tasks that would normally require specialized education. The concern is straightforward: if an AI system can guide an untrained person through complex biological procedures, the barrier to misuse in sensitive areas, including those adjacent to biosecurity, drops significantly.
Anthropic cited the paper as one input in its Responsible Scaling Policy (RSP) version 3.0, the internal framework the company uses to decide whether a model is safe enough to deploy broadly. The RSP sets capability thresholds across domains like biology, cybersecurity, and autonomous behavior. When a model approaches or crosses those thresholds, Anthropic’s policy calls for additional safeguards or restricted access before wider release.
The biology paper, however, is only one piece of a larger evaluation. Anthropic has not published the full set of studies or internal test results that informed the restriction. Whether the biological findings alone triggered the decision, or whether unpublished results in other risk categories played a larger role, remains unclear.
Why the White House got involved
The AP reported that the Wiles-Amodei meeting covered the model’s implications and safety profile, and that Anthropic provided a statement about the encounter. No official transcript or detailed readout has been released by either side. The specific policy outcomes, if any, that came from the conversation have not been disclosed.
The meeting’s significance lies less in what was said than in what it signals. Federal officials treated a private company’s voluntary product restriction as something worth a face-to-face discussion at the chief-of-staff level. That suggests the administration views frontier AI releases not as routine product launches but as events with potential national security dimensions.
It is not clear whether this was a one-time briefing or part of a continuing dialogue between the White House and AI companies about how to handle models that test the boundaries of safe deployment.
What Anthropic has not disclosed
For a company that has built its reputation on transparency and safety-first development, the gaps in public information are notable. Anthropic has not released the full results of its internal evaluations. It has not specified which capability thresholds the new model approached or exceeded. And it has not provided a timeline for when broader access might be granted.
The arXiv paper offers real data, but it addresses only one dimension of risk. Anthropic’s RSP framework also covers cybersecurity, autonomous replication, and persuasion capabilities, and the company has said nothing publicly about how the new model performed in those areas. No U.S. government agency has published an independent safety assessment of the model, either.
Readers and researchers should treat claims about the model’s specific capabilities, whether from Anthropic, competitors, or outside commentators, with caution until more primary data becomes available.
The competitive pressure behind closed doors
Anthropic’s decision does not exist in a vacuum. OpenAI, Google DeepMind, and Meta are all developing frontier models on aggressive timelines. As of May 2026, none of these companies has publicly commented on Anthropic’s restriction or announced similar measures for their own upcoming releases.
The competitive dynamic creates a tension that the industry has discussed for years but never resolved: voluntary restraint by one company could push rivals to release faster in order to capture market share. Whether that pressure is real in this case, or whether Anthropic’s move will encourage more caution across the field, is something only the next round of releases will answer.
What this changes for AI governance
Regardless of what the model can or cannot do, this episode establishes a new pattern. An AI company made a release decision based partly on external empirical research, not just internal benchmarks. That decision triggered a direct conversation with the executive branch. And the result was a restriction that affects researchers, developers, and businesses that depend on access to the most capable AI tools available.
For the broader AI ecosystem, the practical takeaway is that access to frontier models can no longer be assumed. Safety evaluations with ambiguous results may lead to delayed or limited releases, and the factors that determine when restrictions are lifted remain largely outside public view. The gap between what companies know about their own models and what the public is allowed to see is now one of the central tensions in AI policy, and this case has made that gap harder to ignore.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
