Unauthorized users accessed Anthropic's Mythos AI model
Unauthorized users reportedly accessed Anthropic's Mythos AI model, a restricted system being tested with selected companies, by exploiting third-party contractor credentials and using web research techniques to find security gaps. Anthropic stated that the unauthorized access had not compromised confidential information and was investigating the incident.
What happened
A small group of unauthorized users reportedly gained access to Anthropic's Mythos AI model, which was restricted to selected partners during Project Glasswing testing.
The users exploited a third-party contractor's credentials and used web investigation techniques to locate insecure access points, according to Bloomberg reporting.
Users participated in a private online Discord forum and made documented assumptions about the structure of Anthropic's systems, leveraging information from a data breach affecting Mercor, a startup that collaborates with AI companies.
Reported impact
Anthropic stated that as of the reporting, there were no indications that the unauthorized access affected confidential information or the integrity of the Mythos system.
Index analysis
The incident exposed vulnerabilities in access controls for a highly restricted AI system, despite the system being limited to select testing partners.
The breach combined credential exploitation and OSINT techniques, suggesting that even closely guarded AI deployments can face unauthorized access threats.
Governance controls
Anthropic confirmed it was investigating the incident and had been restricting Mythos access to a controlled group of companies including Apple, Amazon, Cisco, Google, Microsoft, and others through Project Glasswing.
The model had been withheld from general public release due to its ability to identify and exploit system vulnerabilities with potential economic, public safety, and national security implications.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.