A small group of unauthorized users gained access to Anthropic PBC's new Mythos artificial intelligence model, which the company claims can facilitate dangerous cyberattacks, UNN reports, citing Bloomberg.
Details
This unauthorized access, previously unreported, highlights the challenge Anthropic faces in fully preventing the spread of its most powerful — and potentially dangerous — technology beyond approved partners. It also raises questions about whether anyone else might be using Mythos without permission and for what purpose.
Claude's creator accidentally declassified their most powerful AI model28.03.26, 11:14
Users employed various tactics to access Mythos. These included leveraging access a person had as an employee of a third-party Anthropic contractor and attempting to use common internet-based information-gathering tools often used by cybersecurity researchers, the source said.
The users are members of a private Discord channel dedicated to finding information about unreleased models, including through bots that search for details published by Anthropic and others on unsecured websites like GitHub.
"We are investigating reports of unauthorized access to Claude Mythos Preview through the environment of one of our third-party vendors," an Anthropic spokesperson said in a statement.
The company stated that it currently has no evidence that the access reported by Bloomberg extended beyond the third-party vendor's environment or that it affected any Anthropic systems.
To date, Anthropic has allowed Apple Inc., Amazon.com Inc., Cisco Systems Inc., and dozens of other organizations to begin testing Mythos. Amazon, a key partner and sponsor of Anthropic, also offers Mythos through its Bedrock platform to a limited list of approved organizations.
In recent days, more financial institutions and government agencies on both sides of the Atlantic are seeking to be included in the list of early testers to protect their systems from malicious actors.
To gain access to Mythos, the group of users made an educated guess about the model's location on the network, based on knowledge of the format Anthropic used for other models, the source said, adding that such details were revealed in a recent data leak from Mercor, an AI training startup working with a number of leading developers.
Importantly, this individual also has permission to access Anthropic models and software used to evaluate the startup's technology. This access was obtained from a company for which they performed contract work evaluating Anthropic's AI models.
The group, as the source indicated, is interested in experimenting with new models rather than harming them. The group did not run cybersecurity-related tests on the Mythos model, preferring instead to perform tasks such as creating simple websites to avoid detection by Anthropic.
This person said the group also has access to many other unreleased Anthropic AI models.