Key Takeaways

  • Anthropic’s Mythos model uncovered vulnerabilities in classified U.S. government systems within hours during Project Glasswing testing.
  • The U.S. administration restricted access to Anthropic’s Fable 5 and Mythos 5 models, citing national security concerns.
  • Industry frameworks from NIST, CISA, and ENISA highlight the dual-use implications of AI-assisted vulnerability discovery.

Anthropic’s involvement in a recent federal cybersecurity exercise has touched off a broader debate about the role of frontier artificial intelligence in national security, raising questions for technology leaders about how quickly these models are evolving. The company’s Mythos model identified vulnerabilities in highly sensitive U.S. government systems within hours during a controlled test, according to a U.S. official who described the results to The Associated Press. The testing took place under Project Glasswing, an initiative that aims to coordinate industry and federal agencies around emerging software and infrastructure risks.

Security leaders focused closely on the context surrounding the model's testing speed. According to reporting from the Bangor Daily News, which documented key details in the days before this disclosure, the model pinpointed weaknesses within hours, faster than teams had anticipated. This acceleration introduces new defensive capabilities while elevating operational risks. A Reuters account, referenced widely across the security community, reinforced that the model’s findings did not mean Mythos could exploit what it discovered within the same compressed timeframe.

Details emerging from a Senate hearing on June 11 added fuel to the policy conversation. A U.S. senator cited remarks from the director of the National Security Agency and U.S. Cyber Command, noting that the tool had penetrated almost all classified systems it was evaluated against. The NSA did not expand on those comments afterward, and Anthropic itself has stayed quiet.

The tension increased when the U.S. administration issued a directive prohibiting foreign nationals from using Anthropic’s newest models, Fable 5 and Mythos 5. Anthropic complied by disabling access for all customers. The company expressed concern that the restrictions were not proportional to the issue it had flagged to federal agencies, especially given how limited access to Mythos already was.

Several cybersecurity executives pushed back on the directive. They argued that limiting access to tools like Mythos could hinder defenders more than adversaries, especially since multiple companies and open-source communities maintain models that offer comparable capabilities. Several industry experts, including leaders from Adobe and Nvidia, argued that removing advanced AI tools from the hands of domestic defenders during a period of accelerated global cyber competition introduces unnecessary risk.

Industry bodies have been mapping similar concerns for the past several years. The National Institute of Standards and Technology (NIST) has repeatedly cautioned that AI-driven vulnerability discovery is a dual-use capability. In its AI Risk Management Framework, NIST outlines how tools that assist with secure code review also carry the potential to support rapid exploit development. The guidance is not intended as a barrier to innovation, but as a reminder that oversight structures need to adapt as quickly as the models themselves.

The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has also highlighted the speed shift. Its recent advisories note that automated reconnaissance and exploit generation are accelerating across both public and private sector networks. CISA has encouraged government programs to adopt secure-by-design approaches in procurement and system development, something many federal agencies are still working into their modernization strategies.

The European Union Agency for Cybersecurity (ENISA) has also taken a broad view on AI-assisted red teaming. ENISA analysts point out that while automated testing tools can raise the baseline of defensive capability, they also lower the barrier for sophisticated attackers. That contrast is central to the ongoing policy dispute in Washington. Anthropic’s internal caution about how the U.S. military might deploy its models intersects awkwardly with federal agencies’ desire to use advanced AI to improve their own cyber readiness.

The situation demonstrates how fast frontier models are moving from experimental to operational relevance. The testing conducted under Project Glasswing was intended to help the government understand the practical implications of these models as part of a broader risk management effort, yet it has also surfaced questions about how government departments will collaborate with vendors under new vetting rules. The executive order signed by the U.S. president earlier this month gives the government up to a month to assess national security risks for advanced AI releases, although participation by developers remains voluntary.

Organizations across critical infrastructure are now evaluating how these developments will impact their own security requirements. Vendors like OpenAI, Google DeepMind, and platforms such as Recorded Future are all testing AI-assisted threat analysis. The combination of increasingly capable models and emerging federal guidance is likely to influence procurement decisions and development roadmaps over the next few years.

While specific details of the model's performance remain classified, accelerated vulnerability discovery carries clear benefits for defenders, especially when paired with disciplined oversight structures. Yet it also challenges the assumptions that have governed federal cyber operations and software testing for decades. The debate between Anthropic and the administration shows that the path to deploying advanced AI in national security settings will require both transparency and negotiation, even among partners who share similar goals.

Whether the directive restricting Fable 5 and Mythos 5 access is temporary or becomes standard practice is not yet clear. AI-assisted security testing has moved from hypothetical to operational faster than many expected, and both government agencies and technology companies are actively adjusting their cyber readiness frameworks in response.