admin The decision allows more than 100 vetted organisations to resume use of the company’s most advanced cyber-focused artificial intelligence model. The approved users are understood to include major enterprises, government-linked agencies and institutions responsible for defending networks in sectors such as energy, finance, healthcare, transport and communications.
Access had been suspended on June 12 after federal authorities raised national security concerns over the model’s capabilities and possible misuse by foreign adversaries. The intervention marked one of the strongest signs yet that Washington is prepared to treat frontier AI systems not merely as software products, but as strategic technologies requiring direct government oversight.
Claude Mythos 5 is designed for advanced cybersecurity work, including vulnerability discovery, threat analysis and defensive code review. Its earlier preview version was made available through Anthropic’s Project Glasswing, a programme aimed at helping trusted partners identify and fix software weaknesses before they can be exploited. The initiative was expanded this month to roughly 150 organisations in more than 15 countries, subject to security checks.
The reinstatement is narrower than Anthropic had sought. Access remains limited to approved U. S. organisations, while broader availability of Claude Fable 5, the company’s widely released high-end model, has not fully resumed. The restrictions have left developers, corporate customers and security researchers uncertain about the rules governing future access to frontier models.
The June 12 suspension affected both Claude Mythos 5 and Claude Fable 5. Anthropic disabled access more broadly after being told to restrict use by foreign nationals, a requirement that proved difficult to enforce in real time across cloud services, corporate teams and API customers. The company then entered talks with federal officials to restore access for organisations considered essential to national cyber defence.
The approved carveout reflects a compromise between competing policy goals. Officials are trying to prevent powerful AI tools from being used to accelerate offensive cyber operations, while security teams argue that the same systems are needed to defend complex networks against state-backed hacking groups and criminal ransomware crews.
Cybersecurity specialists have warned that blanket restrictions can weaken defence if they prevent trusted teams from using frontier models to triage alerts, analyse malware, test code and patch vulnerabilities. Large infrastructure operators face persistent pressure from espionage campaigns, supply-chain compromise and attacks on operational technology systems. AI tools are increasingly being tested as force multipliers for small teams handling large volumes of telemetry and code.
The model’s strengths have also made it a focus of regulatory concern. Advanced AI systems can assist with legitimate vulnerability discovery, but the same capabilities may help malicious actors identify exploitable flaws, automate reconnaissance or adapt attack methods faster than defenders can respond. That dual-use nature has pushed cybersecurity to the centre of AI governance debates.
Anthropic has positioned Mythos as a controlled-access model rather than a mass-market product. Its public materials describe Mythos 5 as available only to a small group of vetted partners, with the aim of expanding access over time. The company has said access decisions are linked to safety requirements, partner vetting and the model’s higher-risk capabilities.
The federal action follows months of intensifying scrutiny of advanced AI systems. Policymakers have been moving beyond chip export controls towards direct supervision of model deployment, especially where systems show strong performance in cybersecurity, biotechnology, software engineering and autonomous agent tasks. The Mythos case is being watched by rivals, cloud providers and enterprise customers as a test of how far the government may go in controlling frontier AI releases.
The episode has also exposed tensions inside the AI industry. Companies want predictable rules before investing heavily in specialised models, while government agencies want visibility into capabilities that could alter national security risks. Customers, meanwhile, want assurance that access to critical tools will not be interrupted without clear process or notice.
