AI development company Anthropic has officially confirmed an accidental leak of information about its most powerful AI model to date, UNN reports with reference to Mashable.
Details
The model, now known as Claude Mythos, was first revealed in a Fortune report. Anthropic confirmed the leak details to the publication.
According to Fortune, the data leak included information about the upcoming release of the Claude Mythos AI model, an "exclusive CEO event," and other internal materials such as PDFs and images. In total, the leak included nearly 3,000 pieces of content that Anthropic had not previously published.
According to Anthropic, the leak was caused by an issue with the content management system (CMS) the company uses. Anthropic did not mark these items as private, so the data was stored in a publicly accessible data repository.
According to the Claude Mythos leak, Anthropic has completed training the model and calls it "the most powerful AI model we have ever developed to date." In a statement provided to Fortune after the leak, an Anthropic spokesperson described Claude Mythos as a "qualitative leap" in AI performance and stated that the model is "the most functional of all we have created to date." Anthropic states that Claude Mythos is currently in the testing phase and available to selected "early access customers."
In addition to the new AI model, the report also mentions a new, yet-to-be-released tier of Anthropic's AI models called Capybara. Currently, Anthropic offers three different tiers for each of its AI models: Opus, Sonnet, and Haiku. Opus is currently the most powerful and expensive tier, while Haiku is the smallest and least expensive. Capybara will be even above Opus, being the largest and most intelligent tier of Anthropic's AI models.
The report also details Anthropic's concerns about the cybersecurity risks posed by Claude Mythos. Apparently, Anthropic believes that Claude Mythos could be used by hackers for cyberattacks. "As we prepare for the release of Claude Capybara, we want to proceed with extreme caution and understand the risks it presents – even beyond what we learn through our own testing," the Anthropic report states. "Specifically, we want to understand the potential short-term risks associated with this model in the field of cybersecurity and share the results to help cybersecurity professionals prepare."
Anthropic described its yet-to-be-released model as "currently significantly ahead of any other AI model in the cybersecurity industry" and stated that the model "anticipates a future wave of models that will be able to exploit vulnerabilities in ways that significantly surpass the efforts of defense professionals." In the leaked blog post, the company stated that it is providing early access to the project to give organizations "a head start in improving the reliability of their codebases against a wave of AI-driven exploits."
US court blocks Trump administration's ban on Anthropic AI in Pentagon saga27.03.26, 18:39