Anthropic announced this week that it has limited the release of its latest model, Mythos, due to its high capability in discovering security exploits in widely used software. Instead of making Mythos publicly available, the lab will share it with major companies and organizations operating critical online infrastructure, like Amazon Web Services and JPMorgan Chase. OpenAI is reportedly considering a similar strategy for its next cybersecurity tool, aiming to enable these large enterprises to preempt bad actors using advanced LLMs for software breaches.
The reference to “exploits” hints that there might be motivations beyond just cybersecurity or model hype behind this strategy. Dan Lahav, CEO of the AI cybersecurity lab Irregular, commented that while AI tools’ discovery of vulnerabilities is important, an exploit’s significance depends on factors like combinability.
Lahav questioned if the findings were meaningfully exploitable alone or as part of a sequence. Anthropic claims that Mythos surpasses its predecessor, Opus, in exploiting vulnerabilities. Yet, it’s unclear if Mythos is the ultimate cybersecurity model. Aisle, an AI cybersecurity startup, managed to replicate much of Mythos’s reported achievements using smaller, open-weight models, suggesting that no single deep learning model fits all cybersecurity needs.
Opus was already considered transformative for cybersecurity, and there may be another motive for restricting distribution to large organizations: securing enterprise contracts while preventing competitors from using distillation to replicate models easily and cheaply.
David Crawshaw, CEO of exe.dev, noted that high-end models are now restricted by enterprise agreements, limiting access for smaller labs. He suggested this strategy sustains enterprise revenue by relegating distillation companies to a lower tier, as new top models remain enterprise-exclusive upon widespread release.
This situation reflects trends in the AI industry: A competition between labs developing major models and companies like Aisle that advocate for varied approaches using open-source LLMs, often from China and allegedly developed via distillation, for economic benefit.
Frontier labs are increasingly opposing distillation; Anthropic has openly disclosed attempts by Chinese firms to copy its models, with Anthropic, Google, and OpenAI collaborating to identify and block distillers as per a Bloomberg report. Distillation threatens frontier labs’ business models by negating the advantage of large-scale capital investment. Blocking distillation is beneficial, while selective model release allows labs to differentiate enterprise offerings as this category becomes central for profitability.
It remains uncertain if Mythos or future models will truly endanger internet security, but gradual technology deployment is prudent. Anthropic did not respond regarding potential distillation concerns at the time of writing, but the company might have found a smart way to protect both the internet and its financial interests.