Contractor Granted Unauthorized Users Access
A group of unauthorized individuals has gained access to Claude Mythos Preview – Anthropic's most advanced and restricted AI model. This was reported by Bloomberg, as cited by The Verge. According to the report, the group utilized a combination of access to a named third-party contractor for Anthropic and what Bloomberg describes as "commonly used internet investigation tools."
Access occurred via a private online forum, and it is currently unclear exactly how many people gained access to the model, or what they might have used it for. Anthropic has not publicly commented on the extent of the incident.
What Makes Mythos So Dangerous?
Claude Mythos Preview – internally codenamed "Capybara" – is an AI model Anthropic has explicitly withheld from public release due to its offensive cybersecurity potential. According to research related to the case, the model's capabilities are significantly beyond what has been seen in previous AI systems.
Sarosh Nagar, technical project lead in the office of former Google CEO Eric Schmidt, characterizes Mythos' ability to automatically chain vulnerabilities as "significantly beyond what previous models have been able to do," according to research associated with the case.
Model Announced After a Leak
Anthropic's official announcement of Claude Mythos Preview came on April 7, 2026 – but by then, the model had already become known via an accidental leak of a draft blog post in late March 2026, caused by a misconfiguration in the company's publishing tool.
Project Glasswing – Defensive Use with Tech Giants
To manage the risks associated with the model, Anthropic has launched "Project Glasswing," an initiative where Mythos will primarily be used for defensive security research. According to the research material, the consortium includes companies such as Amazon, Apple, Google, and Microsoft, in addition to financial institutions like JPMorgan Chase.
The goal is to use Mythos to secure critical software infrastructure – that is, to leverage the same offensive capabilities to find and patch security holes before other actors can exploit them.
Previous Misuse of Anthropic Systems
This incident is not the first time Anthropic's AI systems have been involved in unauthorized activity. According to the research material, a Chinese state-sponsored actor previously used Anthropic's Claude Code tool to automate cyberattacks against approximately 30 companies and government agencies. These operations included reconnaissance, scanning for vulnerable systems, generating phishing content, and data exfiltration – largely without human involvement.
The fact that Mythos – a significantly more powerful system – has now fallen into the hands of unauthorized users will likely increase pressure on Anthropic and the industry at large regarding access control and the responsible handling of advanced AI models.
Source material: The Verge / Bloomberg.