Unauthorized users gained access to Anthropic's AI model Mythos, raising security concerns. The company is investigating the breach.
A group of unauthorized individuals has accessed Anthropic's newly launched AI model, Mythos, according to a report by Bloomberg based on internal documents.
Members of a private online forum reportedly utilized the neural network on its release day and have continued to do so regularly.
Anthropic describes Mythos as capable of identifying and exploiting vulnerabilities across major operating systems and web browsers. Consequently, the company has limited access to a select group of software developers.
The breach involved several tactics, including using credentials from a third-party contractor, guessing the model's URL based on patterns from other company systems, and leveraging information from a data leak at the startup Mercor.
A source from Bloomberg indicated that the group plans to experiment with the model without malicious intent. They also reportedly have access to other unreleased neural networks developed by Anthropic.
"We are investigating reports of unauthorized access to the Claude Mythos Preview through one of the environments of third-party providers," a company spokesperson stated.
This incident underscores the difficulties in controlling potentially hazardous technologies and raises concerns about who else might gain access to Mythos and their intentions.
Capabilities of Mythos
Mozilla noted in a blog post that an early version of Mythos detected 271 vulnerabilities in the Firefox browser during internal testing, all of which have since been addressed.
This illustrates how advanced AI can analyze extensive codebases to uncover weaknesses that typically require extensive scrutiny by cybersecurity professionals.
Previously, Mozilla had tested another Anthropic model that identified 22 vulnerabilities in an earlier Firefox version. Despite these findings, the company asserts that achieving complete security is an unrealistic goal.
"Some commentators believe future AI models will uncover entirely new forms of vulnerabilities. We do not share that view," the firm noted.
In April, reports emerged that the U.S. National Security Agency has been utilizing Mythos, despite the ongoing tensions between the startup and the Pentagon. Additionally, in March, Anthropic confirmed a leak of part of its AI programming tool, Claude Code, and later accidentally deleted thousands of GitHub repositories while attempting to resolve the issue.
