Anthropic’s top brass arrived in Washington, D.C., on Monday to meet with senior officials from the White House and the Commerce Department, hoping to defuse a growing clash over the company’s newest AI system, Claude Fable 5. The talks, attended by co‑founder and chief compute officer Tom Brown, head of external affairs Sarah Heck, frontier red‑team leader Logan Graham, and senior security researcher Nicholas Carlini, concluded without the administration lifting the export controls imposed just a week earlier.
U.S. officials remain convinced that the model’s safeguards can be stripped away, granting users access to the more potent capabilities of Anthropic’s Mythos‑class models. The concern centers on a potential “jailbreak” that would let users bypass guardrails protecting sensitive domains such as cybersecurity, biology and chemistry. While Anthropic argues that the risk is exaggerated, the Commerce Department’s stance reflects a broader uncertainty about how advanced foundation models might be weaponized.
During the meetings, representatives from the Center for AI Standards and Innovation and the Office of the National Cyber Director, led by Sean Cairncross, participated alongside Commerce Secretary Howard Lutnick, who joined by conference call from the G7 summit in Evian, France. A White House spokesperson declined to comment on the outcome, and the administration’s next steps remain unclear.
Anthropic’s spokesperson told WIRED that both sides are working quickly to resolve the issue, but no timeline was offered. The Commerce Department signaled a willingness to restore consumer access to Fable 5 if Anthropic can fully address the jailbreak concerns, though no concrete roadmap was disclosed.
The dispute arrives amid a broader political pressure cooker for Anthropic. The company is already locked in a protracted fight with the Pentagon over whether its models should be used for certain military applications. Last week, Amazon CEO Andy Jassy contacted Treasury Secretary Scott Bessent about the alleged vulnerabilities, prompting the administration to task the NSA with a rapid review. The NSA concluded that it was indeed possible to strip away Fable 5’s guardrails, which led to the swift imposition of export controls.
Amazon, a major investor in Anthropic, declined to elaborate on its internal discussions, stating only that it routinely advises governments on potential security risks for its cloud customers. The involvement of such a high‑profile backer has heightened scrutiny of the situation, with some investors worried that Anthropic is being singled out while competitors may not face comparable actions.
Security researchers pushed back against the government’s characterization of the risk. An open letter signed by dozens of experts argued that the export control action unfairly removes the best models from defenders, creates market uncertainty, and threatens U.S. AI leadership without solid justification. They noted that while Mythos‑class models excel at finding and weaponizing exploits, they are not uniquely capable of these tasks; many security professionals already use other foundation and open‑source models for red‑team work.
Anthropic’s own blog post earlier this week suggested that the model’s safeguards are sufficient for public release, despite acknowledging that a fully successful jailbreak would effectively turn Fable 5 into an unrestricted Mythos model. Independent analysts, such as Luta Security founder Katie Moussouris, concluded that the reported vulnerabilities did not constitute a true jailbreak, describing the guardrails as “speed bumps” rather than hard security boundaries.
Looking ahead, the episode raises questions for the entire AI industry about how to navigate government oversight when releasing powerful models. Leaders of other labs, including Cohere CEO Aidan Gomez, warned that the White House’s willingness to impose export controls signals a new reality: AI developers may need to provide early access to advanced models and maintain proactive communication with regulators to avoid similar clashes.
For now, Anthropic’s Claude Fable 5 remains offline for consumer use, and the company continues to negotiate with U.S. officials. The outcome will likely shape not only Anthropic’s roadmap but also the broader framework governing the deployment of high‑risk AI technologies in the United States.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.