Artificial intelligence firm Anthropic has limited access to its latest model, Mythos, citing concerns over its advanced ability to identify and exploit software vulnerabilities. Rather than releasing the model publicly, the company said it will make it available only to select organisations managing critical infrastructure, including firms such as Amazon Web Services and JPMorgan Chase.
The move reflects growing caution among leading AI developers about the dual use nature of powerful large language models, particularly in cybersecurity. Reports indicate that OpenAI is considering a similar approach for its upcoming cybersecurity tools, with the aim of enabling major enterprises to strengthen defences before such technologies can be exploited by malicious actors.
However, some industry experts have questioned whether security concerns are the sole motivation behind the restricted rollout. Dan Lahav, whose firm focuses on AI driven cybersecurity, noted that the real world impact of vulnerabilities depends on how they can be combined and deployed, suggesting that no single model represents a complete breakthrough. Other startups have also claimed that similar capabilities can be replicated using smaller or open source models, raising doubts about Mythos’ uniqueness.
Analysts say the strategy may also serve commercial interests, particularly in limiting access to top tier models to large enterprise clients while curbing the practice of model distillation, a technique used to replicate advanced AI systems at lower cost. As competition intensifies between major labs and smaller players leveraging open models, the selective release of high end systems like Mythos highlights a broader shift in the AI industry, where control over access is becoming as critical as the technology itself.