Anthropic recently announced a strategic limitation on the public release of its latest model, **Mythos**, due to its advanced capabilities in identifying security vulnerabilities in widely-used software. Rather than making Mythos available to the general public, the AI research lab plans to share it exclusively with a select group of major corporations and organizations that manage critical online infrastructure, including Amazon Web Services and JPMorgan Chase. This approach is designed to enable these enterprises to stay ahead of potential threats posed by malicious actors utilizing sophisticated large language models (LLMs).
However, the choice of a restricted rollout raises questions about whether the motivation goes beyond merely enhancing cybersecurity. Dan Lahav, CEO of the AI cybersecurity firm **Irregular**, emphasized that while AI tools can uncover vulnerabilities, the significance of these weaknesses to attackers is contingent on various factors, including their potential for exploitation in conjunction with other vulnerabilities.
Anthropic claims that Mythos can exploit vulnerabilities more effectively than its predecessor, **Opus**. Yet, it's unclear if Mythos represents the pinnacle of cybersecurity solutions. The startup **Aisle** has asserted that it can replicate many of Mythos's capabilities using smaller, open-weight models, suggesting that no single deep learning model can comprehensively address cybersecurity challenges.
Moreover, limiting access to advanced models may serve a dual purpose for frontier labs: it fosters lucrative enterprise contracts and complicates competitors' efforts to replicate their models through distillation--a method that allows new LLMs to be trained more affordably using established models.
David Crawshaw, a software engineer and CEO of **exe.dev**, noted that the current trend appears to prioritize enterprise agreements, effectively sidelining smaller labs and ensuring a steady flow of revenue from larger organizations. This dynamic highlights the ongoing competition within the AI ecosystem, where frontier labs are racing to develop the most advanced models, while companies like Aisle leverage multiple models and open-source LLMs to gain a competitive edge.
This year, frontier labs have adopted a stricter stance on distillation, with Anthropic publicly addressing attempts by Chinese firms to replicate its models. In collaboration with Google and OpenAI, Anthropic is actively working to identify and thwart distillation efforts, recognizing it as a significant threat to their business models.
As the rollout of Mythos progresses, the implications for internet security remain uncertain. A measured introduction of this technology could pave the way for responsible advancements in cybersecurity. While Anthropic has not commented on whether distillation concerns influenced their decision, it appears they may have devised a strategic method to safeguard both the internet and their commercial interests.