.png){kind=logo}
Anthropic and the US government are discussing a new framework for assessing security flaws in advanced artificial intelligence models. The talks follow federal restrictions on foreign access to Claude Fable 5 and Mythos 5.
The proposed standards would help officials and AI companies measure the seriousness of future vulnerabilities. They could also guide decisions on model access, export limits and other government action.
White House officials and Anthropic representatives are reportedly working on shared rules for reviewing security weaknesses. The discussions focus on when a flaw poses a national security risk and how quickly authorities should respond.
The framework could examine whether users can bypass a model’s safeguards, how easily they can repeat the method and what the system can produce after the controls fail. Officials also want to consider whether a weakness can cause real-world harm.
Anthropic’s Head of Public Policy Sarah Heck and co-founder Tom Brown are reportedly leading the company’s involvement. Senior researchers and safety specialists also met Commerce Department officials in Washington this week.
Neither side has announced a final agreement. Therefore, it is unclear when the framework could take effect or whether it would apply to models from other US AI companies.
Anthropic launched Claude Fable 5 and Mythos 5 before the US government issued an export control directive. The government ordered the company to stop foreign nationals from accessing the advanced systems, citing national security concerns.
The action followed concerns about a reported ‘jailbreak’ vulnerability. A jailbreak uses prompts or other methods to bypass an AI model’s built-in safety controls. Such access may allow a system to provide information that its safeguards would normally block.
Anthropic disabled both models for users worldwide after receiving the order. The company said it could not reliably separate every permitted user from people covered by the restrictions across its services and cloud platforms.
Anthropic also questioned the standard used in the decision. The company stated, “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”
President Donald Trump addressed the discussions at the G7 summit in Évian-les-Bains, France. When reporters asked about the talks with Anthropic, Trump said negotiations were ‘going fine.’
Trump gave no further details about any proposed settlement or possible return of the models. A White House spokesperson also cited “national security concerns” but did not explain the specific risks under discussion.
The summit meeting marked Trump’s first public meeting with Anthropic CEO Dario Amodei since the export order. Several technology executives attended discussions with G7 leaders, including OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis.
Meanwhile, some early Mythos users reportedly retained access through a controlled testing program. The conditions governing that access have not been publicly explained, creating questions about which organisations qualify for exemptions.
Amodei and Hassabis reportedly called for a US-led international coalition during the closed-door G7 meeting. The proposed group would develop shared AI rules and security standards among participating countries.
Amodei discussed controlled access to frontier models and cooperation on cyber, biological and intelligence risks. He also supported closer coordination on advanced chips and other critical technology, according to people familiar with the meeting.
Altman separately called for an international forum that could establish accepted testing standards and provide independent reviews of AI capabilities and risks. No binding international agreement was announced after the discussions.
For now, Anthropic and US officials have not confirmed when access to Fable 5 and Mythos 5 may return. The ongoing talks are focused on defining clearer tests before similar security disputes emerge around future AI releases.
Also Read: Anthropic’s Claude Mythos Briefly Surfaces Online After Restricted Access Claims
