Anthropic's Mythos models shut down for two weeks after White House AI security intervention

Anthropic's Mythos Models Halted Following White House Security Intervention

Anthropic shut down all access to its Mythos models for approximately two weeks after an unprecedented White House national security intervention. The NSA flagged the models' ability to hack into systems during a red-team exercise, prompting the administration to overhaul its light-touch approach to AI oversight.

The intervention represents a significant shift in how the Trump administration handles frontier AI capabilities. Anthropic had previously worked with the administration on a limited rollout of an earlier Mythos version, but the hacking demonstration triggered a broader crackdown. Access appears to be in the process of being restored, though a formal approval process is now in place for release.

OpenAI followed a similar trajectory with its GPT-5.5 Cyber model, which demonstrated the ability to discover software vulnerabilities usable in cyberattacks. The company said it is limiting access to its newest models after discussions with the White House, but pushed back on the precedent. "We don't believe this kind of government access process should become the long term default," OpenAI said in a blog post. "It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

The company indicated it hopes to make GPT-5.6 generally available in the coming weeks and framed the current government approval process as a transition period while Trump's recent executive order on model oversight is implemented.

The distillation problem

The case-by-case intervention creates a regulatory tension that may prove unsolvable. As the open-source frontier advances—driven by distillation of frontier models through consumer API access—keeping closed-source models locked behind government approval becomes less tenable. Groups can create networks of thousands of accounts to extract advanced capabilities from widely available models, narrowing the gap between frontier and open-source systems.

Bill Gurley suggested one potential path forward: if these models become sophisticated enough to detect distillation attacks, companies could shift from a whitelist approach to a dynamic ban list, using the models themselves to identify and block suspicious behavior patterns in real time. This would theoretically keep models available to legitimate users while creating a moving target for bad actors. The approach would require constant enforcement but would avoid the blanket restrictions that safety researchers have criticized as both ineffective and overly concentrated in government hands.

The core tension remains unresolved: regulatory solutions that maintain open access appear structurally incompatible with containing frontier cybersecurity and biosecurity capabilities. Keeping models restricted requires either government-administered approval processes that concentrate power in the state, or preventing distillation through technical or legal means—neither of which has proven workable at scale.