Anthropic announced this week that it will not release its latest AI model, called Mythos, to the general public. The company says the model is so capable at discovering security vulnerabilities in widely used software that a broad release could pose serious risks.
Instead, Mythos will be shared with a select group of major corporations and organizations that manage critical online infrastructure, including Amazon Web Services and JPMorgan Chase. OpenAI is reportedly considering a similar limited-release approach for its own upcoming cybersecurity tool. The stated goal is to give these large enterprises a head start in securing their systems before malicious actors can exploit AI-powered hacking capabilities.
How Dangerous Is Mythos, Really?
Not everyone is convinced by Anthropic's framing. Dan Lahav, CEO of AI cybersecurity firm Irregular, noted that the real significance of any discovered vulnerability depends on multiple factors, including whether it can be meaningfully exploited on its own or as part of a broader attack chain.
While Anthropic claims Mythos can exploit vulnerabilities far more effectively than its predecessor, Opus, cybersecurity startup Aisle says it replicated much of what Mythos achieved using smaller, open-weight models. Aisle's team argues this demonstrates that no single large model dominates cybersecurity — results depend heavily on the specific task.
The Real Play: Enterprise Contracts and Anti-Distillation
According to TechCrunch's analysis, the limited release may serve a dual purpose that goes well beyond safety. Restricting access to top-tier models creates a flywheel for lucrative enterprise contracts while simultaneously making it harder for competitors to copy these models through distillation — a technique where frontier models are used to cheaply train new, smaller ones.
Software engineer and exe.dev CEO David Crawshaw argued on social media that this is essentially marketing cover for gating top-end models behind enterprise agreements, keeping them out of reach of smaller labs that would distill them. He predicted that by the time Mythos reaches ordinary users, a newer model will already be locked behind enterprise-only access, creating a perpetual treadmill that keeps enterprise revenue flowing.
A Broader Industry Battle
This move reflects a deepening rift in the AI ecosystem. On one side are frontier labs like Anthropic, Google, and OpenAI, which spend enormous sums building the largest and most capable models. On the other are companies that rely on open-source and distilled models — often originating from Chinese firms — as a cheaper path to competitive AI capabilities.
Frontier labs have taken an increasingly aggressive stance against distillation this year. Anthropic has publicly accused Chinese firms of attempting to copy its models, and according to Bloomberg, Anthropic, Google, and OpenAI have formed a joint effort to identify and block distillers.
Distillation is a fundamental threat to frontier labs because it erases the competitive advantage that comes from massive capital investment in scaling. A selective release strategy addresses both concerns at once: it protects models from being copied while simultaneously creating premium enterprise products.
The Bottom Line
A cautious rollout of a powerful cybersecurity model is undoubtedly a responsible approach. But industry observers say safety is only part of the equation. Anthropic did not respond to questions about whether the decision also relates to distillation concerns. Still, it appears the company may have found an elegant way to protect the internet and its bottom line at the same time.
