Anthropic has raised concerns about its AI model’s cybersecurity risks, as revealed in a leaked post. The internal leak uncovered company information, including Anthropic’s latest powerful AI model, Claude Mythos. The accidental leak was confirmed by the company and reported by Fortune. The leak detailed plans for Claude Mythos, an exclusive CEO event, and other internal documents, totaling nearly 3,000 unpublished assets.
The leak resulted from an issue with Anthropic’s content management system, which failed to keep certain data private, storing it in a publicly accessible data lake. Some of the data were unused from previous announcements, while other parts detailed the CEO event and internal information.
According to a leaked draft, Claude Mythos is described by Anthropic as their most powerful AI model, a significant milestone in AI capabilities. The model is in trial with selected early access customers. The leak also mentioned another AI model tier named Capybara, positioned above the current highest tier, Opus. Capybara is described as Anthropic’s largest and most intelligent AI model yet.
Anthropic has noted the cybersecurity risks associated with Claude Mythos, cautioning that the model could potentially aid hackers in executing cyberattacks. In its leaked draft, Anthropic expressed a commitment to understanding the model’s cybersecurity risks and sharing insights to fortify defenses. The company described the unreleased model as being significantly advanced in cyber capabilities, foreseeing an influx of models capable of exploiting vulnerabilities at a pace outmatching defenders’ efforts.
Early access provided to organizations aims to help them enhance the security of their infrastructures ahead of potential AI-driven exploits.