Anthropic Restricts Mythos: Security or Business Strategy?

Anthropic Restricts Access to Mythos for Security Reasons

This week, Anthropic made the decision to limit the distribution of its latest artificial intelligence model, named Mythos. The reason cited is this model's exceptional ability to identify security vulnerabilities in software used by millions of users worldwide, which could potentially be exploited by malicious actors. Rather than making Mythos available to the general public, Anthropic has chosen to share it only with a select group of influential companies and organizations, such as Amazon Web Services and JPMorgan Chase.

A Strategy Shared by Other AI Giants

OpenAI, another major player in the field of artificial intelligence, is considering a similar approach for its upcoming cybersecurity tool. The common goal is to enable these large companies to stay ahead of cybercriminals who might use sophisticated language models to compromise secure systems. However, the term "e-word" in this strategy suggests that there may be more at stake than just cybersecurity.

Doubts About Mythos's Effectiveness

Dan Lahav, CEO of the AI cybersecurity company Irregular, expressed reservations about the actual impact of Mythos during an interview with TechCrunch in March, even before Mythos was released. According to him, while the detection of vulnerabilities by AI tools is crucial, the significance of these vulnerabilities for an attacker depends on many factors, including their individual or combined exploitability. Lahav questions whether Mythos has truly discovered significantly exploitable vulnerabilities.

Comparison with Other Models

Anthropic claims that Mythos surpasses its predecessor, Opus, in vulnerability detection. Opus was already considered a game-changer for cybersecurity. However, the startup Aisle has stated that it has managed to replicate much of Mythos's capabilities using smaller, more accessible models. This suggests that cybersecurity does not rely on a single model but on adaptation to specific tasks.

A Business Strategy Behind the Limitation

The decision to limit Mythos to large organizations may also be driven by commercial reasons. By restricting access, Anthropic creates leverage to secure lucrative contracts with companies while complicating the task for competitors who might wish to copy their models through distillation. This technique allows for the training of new language models at a lower cost by drawing inspiration from existing models.

Distillation, a Threat to Leading Labs

David Crawshaw, CEO of the startup exe.dev, suggested that this limitation is partly a way to protect high-end models from distillation attempts. According to him, leading models are now controlled by corporate agreements, making access difficult for smaller labs. This strategy helps maintain the revenue of large companies while relegating distillation players to the background.

A Response to International Competition

This year, leading labs have adopted a stricter stance on distillation, particularly in response to attempts by some Chinese companies to copy their models. Three leading labs—Anthropic, Google, and OpenAI—have partnered to identify and block these distillers, according to a Bloomberg report. Distillation threatens their business model by reducing the advantage of the massive investments required to develop these technologies.

A Cautious Approach for the Future

The question of whether Mythos or other similar models pose a threat to Internet security remains open. However, a controlled deployment of these technologies seems to be a responsible approach. Anthropic has not responded to inquiries about a potential link between this decision and concerns related to distillation, but it is possible that the company has found a clever way to protect both the Internet and its financial interests.