Anthropic's Mythos: Is its limited release for public good, or self-interest?

Anthropic announced this week that it has restricted the release of its latest model, dubbed Mythos, due to its advanced capability in identifying security vulnerabilities within software systems relied upon by users globally.

Instead of making Mythos publicly available, the leading AI research lab plans to share it exclusively with a select group of major corporations and organizations that operate critical online infrastructure, spanning from Amazon Web Services to JPMorgan Chase. OpenAI is reportedly considering a similar strategy for its next cybersecurity tool. The stated rationale behind this approach is to enable these large enterprises to proactively counter malicious actors who might leverage sophisticated Large Language Models (LLMs) to penetrate secure software.

However, the prominent focus on "enterprise" in this release strategy suggests motivations potentially extending beyond mere cybersecurity concerns or the promotion of model capabilities.

Dan Lahav, CEO of the AI cybersecurity lab Irregular, told TechCrunch in March, prior to Mythos's release, that while the discovery of vulnerabilities by AI tools is significant, the specific value of any weakness to an attacker depends on numerous factors, including how they can be exploited in combination.

"The question I always have in my mind," Lahav stated, "is did they find something that is exploitable in a very meaningful way, whether individually, or as part of a chain?"

Anthropic asserts that Mythos demonstrates a far greater ability to exploit vulnerabilities than its previous model, Opus. Yet, it remains unclear whether Mythos truly represents the ultimate solution in cybersecurity models. Aisle, an AI cybersecurity startup, reported successfully replicating much of what Anthropic claims Mythos accomplished using smaller, open-weight models. Aisle's team argues that these results indicate there isn't a single definitive deep learning model for cybersecurity, but rather effectiveness is contingent on the specific task at hand.

Given that Opus was already regarded as a significant advancement for cybersecurity, another potential reason for frontier labs to limit their releases to large organizations emerges: this strategy creates a self-reinforcing cycle for substantial enterprise contracts, while simultaneously making it more challenging for competitors to replicate their models using distillation—a technique that leverages frontier models to train new LLMs cost-effectively.

"This is marketing cover for the fact that top-end models are now gated by enterprise agreements and no longer available to small labs to distill," David Crawshaw, a software engineer and CEO of the startup exe.dev, suggested in a social media post. Crawshaw added, "By the time you and I can use Mythos, there will be a new top-end rev that is enterprise only. That treadmill helps keep the enterprise dollars flowing (which is most of the dollars) by relegating distillation companies to second rank."

This analysis aligns with observable trends within the AI ecosystem: a competitive dynamic between frontier labs developing the largest and most capable models, and companies like Aisle, which rely on multiple models and perceive open-source LLMs—frequently originating from China and often purportedly developed through distillation—as a viable path to economic advantage.

Frontier labs have adopted a more stringent stance on distillation this year. Anthropic has publicly revealed what it claims are attempts by Chinese firms to copy its models, and a Bloomberg report indicates that three leading labs—Anthropic, Google, and OpenAI—have collaborated to identify and block distillers. Distillation poses a threat to the business model of frontier labs because it negates the advantages conveyed by investing vast amounts of capital to scale. Therefore, blocking distillation is already a worthwhile endeavor, but the selective release approach also provides these labs with a method to differentiate their enterprise offerings as this category becomes crucial for profitable deployment.

The true extent to which Mythos or any new model might genuinely threaten internet security remains to be seen, and a carefully managed rollout of such technology represents a responsible way forward.

Anthropic did not respond to inquiries regarding whether its decision was also influenced by concerns over distillation by the time of publication. Nevertheless, the company may have devised an ingenious strategy that simultaneously safeguards the internet and enhances its financial performance.