Is Anthropic limiting the release of Mythos to protect the internet — or Anthropic?

This week, Anthropic announced a strategic limitation on the release of its latest model, Mythos, citing its advanced capabilities in identifying security vulnerabilities in widely-used software. Instead of making Mythos available to the general public, the company plans to share it exclusively with major enterprises and organizations that manage critical online infrastructure, such as Amazon Web Services and JPMorgan Chase. The rationale behind this selective approach appears to be a proactive measure against potential threats. By equipping these large entities with Mythos, Anthropic aims to help them stay one step ahead of malicious actors who might exploit the power of sophisticated language models to breach secure systems. This strategy aligns with reports that OpenAI is considering a similar method for its forthcoming cybersecurity solutions. However, the decision raises questions about whether the focus is solely on cybersecurity or if there are underlying business motives at play. Dan Lahav, CEO of the AI cybersecurity firm Irregular, noted earlier this year that the significance of vulnerabilities discovered by AI tools can vary greatly, depending on how they can be utilized in different scenarios. He emphasized the importance of understanding whether any identified flaws have substantial exploitable value. Anthropic claims that Mythos can uncover vulnerabilities with greater efficiency than its predecessor, Opus. Yet, the true effectiveness of Mythos in the cybersecurity landscape remains to be fully determined. In fact, Aisle, an AI cybersecurity startup, has reported replicating much of Mythos's capabilities using smaller, open-weight models, suggesting that a singular deep learning model may not be the definitive answer for cybersecurity challenges. The potential for enterprise contracts is another reason frontier labs may choose to restrict access to their models. By keeping high-end models like Mythos from smaller labs, these companies may enhance their business relationships with larger corporations, while complicating competitors' ability to replicate their innovations through distillation—a process that allows for the training of new language models at a lower cost. David Crawshaw, CEO of exe.dev, pointed out that this approach effectively bars smaller labs from accessing top-tier models, subsequently limiting their competitive edge. As the AI ecosystem evolves, a clear divide is forming between frontier labs focused on developing expansive, powerful models, and companies like Aisle that leverage multiple models and open-source alternatives. This year, frontier labs have intensified their stance against distillation, with Anthropic revealing attempts by Chinese firms to replicate their models. In response, leading labs, including Anthropic, Google, and OpenAI, have united to combat these distillation efforts, recognizing the threat they pose to their business models. While the ultimate impact of Mythos on internet security is yet to be determined, a cautious deployment strategy could be a prudent way forward. Anthropic has not commented on whether its decision is influenced by distillation concerns, but their approach may serve to protect both the internet's integrity and the company's financial interests.