An unintentional leak has been officially acknowledged by AI firm Anthropic concerning its most advanced AI model to date.
The model, now referred to as “Claude Mythos,” was initially revealed in a report by Fortune. Anthropic has since verified the specifics of the leak to the publication.
The leak included information regarding the impending launch of the Claude Mythos AI model, an “exclusive CEO event,” and various internal materials such as PDFs and images, as stated by Fortune. In total, the leak revealed nearly 3,000 assets that Anthropic had not previously made public.
Claude Mythos leak insights
As per Anthropic, the leak resulted from a problem with the content management system (CMS) employed by the company. The data was uploaded to the CMS by Anthropic itself. However, the company neglected to designate the items as private, allowing the data to be stored in a publicly accessible data repository.
Some of the leaked data consisted of unused assets from previous announcements by Anthropic. Other leaked information included internal company details such as employee data or information about an invite-only CEO event.
Additionally, found within the leaked information was insights regarding Claude Mythos from an unpublished blog post draft.
Claude Mythos and Capybara
According to the leaked draft concerning Claude Mythos, Anthropic has finalized the training of the model and describes it as “by far the most powerful AI model we have ever created.”
In a statement issued to Fortune following the leak, an Anthropic representative characterized Claude Mythos as an AI performance “step change” and stated that the model is “the most capable we’ve developed to this point.” Anthropic indicates that Claude Mythos is currently in the trial phase and available to select “early access customers.”
Besides the new AI model, the leaked post also references a new, unreleased AI model tier from Anthropic known as Capybara.
Anthropic presently provides three different tier levels for each of its AI models: Opus, Sonnet, and Haiku. Opus is currently its most potent and most costly tier, while Haiku is the smallest and least expensive. Capybara would be positioned even above Opus as Anthropic’s largest and most intelligent AI model tier.
Capybara cybersecurity concerns
The leaked post furthermore outlines Anthropic’s worries regarding the cybersecurity threats that Claude Mythos introduces. It seems that Anthropic fears that Claude Mythos could be exploited by hackers to conduct cyberattacks.
“In preparation for the release of Claude Capybara, we aim to proceed with heightened caution and comprehend the risks it entails — even beyond what we learn through our own testing,” states the leaked Anthropic document. “Specifically, we intend to grasp the model’s potential immediate risks in the area of cybersecurity — and share the findings to assist cyber defenders in their preparations.”
Anthropic described its unreleased model as “currently far ahead of any other AI model in cyber capabilities” and claimed that the model “heralds an upcoming surge of models that can exploit vulnerabilities in ways that significantly surpass the defenses’ efforts.”
The company noted in the leaked draft blog that it was offering early access to organizations to provide them “a head start in enhancing the robustness of their codebases against the forthcoming surge of AI-driven exploits.”