Claude Mythos leak: model Anthropic called 'too dangerous for the public' reaches outsiders

Anthropic failed to keep access to Claude Mythos under control. According to Bloomberg, in a report published on April 21, 2026, a "small group of unauthorized users" gained access to a system the company itself had spent weeks calling too dangerous for a public release.

How the model fell into the wrong hands

According to Bloomberg, unauthorized access to Mythos has persisted since the day Anthropic first announced plans to open the model to a limited group of companies for testing under Project Glasswing, a cybersecurity initiative. Mythos's very existence had already come to light earlier through a leak, which makes this the second time in a short span that information about the model has escaped the company against its will.

• The leak revealing Mythos's existence came before the model's official announcement
• Anthropic announced Project Glasswing, a plan to test Mythos with a limited group of partners
• According to Bloomberg on April 21, 2026, a "small group of unauthorized users" gained access
• Anthropic confirmed that it is investigating the incident

Why this is awkward for Anthropic

For weeks, Anthropic insisted that Claude Mythos is so strong at cybersecurity tasks that a public release is simply dangerous, while at the same time preparing closed test access for select partners. The Verge calls the episode humiliating for a company that has built its brand on taking AI safety seriously and constantly stresses the cybersecurity capabilities of its models. Anthropic said it is investigating the incident, but it has not disclosed the scale of the breach or how exactly outsiders obtained access.

A "small group of unauthorized users" gained access to

Mythos, according to Bloomberg's report of April 21, 2026.

What we know about Project Glasswing

Project Glasswing is a program announced by Anthropic under which the company planned to give a limited group of companies access to Claude Mythos to test the model's cybersecurity capabilities. Mythos's existence became public ahead of schedule, through a leak rather than an official Anthropic announcement, which already cast a shadow over the release preparations.

What this means

The Mythos episode shows that the tighter a company controls access to a model, and the louder it warns about the model's danger, the more damaging any leak looks. Anthropic now faces a gap between its actual access practices and its public rhetoric about safety. The company that promised to hand Mythos only to select partners under supervision must now explain how people without permission ended up among those with access.

Frequently asked questions

What is Claude Mythos?

It is an Anthropic model that the company publicly called too dangerous for a broad release for weeks because of its cybersecurity capabilities. Officially, it was to be tested only with a limited group of partners.

What is Project Glasswing?

It is an initiative announced by Anthropic to give a limited group of companies access to Claude Mythos for testing cybersecurity scenarios.

Is Anthropic investigating what happened?

Yes. The company confirmed it is investigating the unauthorized access incident, but it has not yet disclosed details, such as how many people gained access or how.