Anthropic's secretive cyber AI Mythos leaks to unauthorized users

Anthropic built an AI tool so dangerous that it restricted access to only a handful of enterprise giants. Naturally, a bunch of Discord users were playing with it that very same day.

The model in question is Claude Mythos Preview, a cybersecurity tool designed to defend corporate infrastructure. Anthropic itself cautioned that if Mythos fell into the wrong hands, it could easily become a potent weapon for launching dangerous cyberattacks. To keep it locked down, Anthropic launched Project Glasswing, an exclusive preview limited to a tiny group of partners, including Apple.

According to a report from Bloomberg, that lockdown didn't last very long.

Finding the back door through vendor pipes

A private online group managed to bypass the restrictions and tap into Mythos through a third-party vendor environment. They gained access on the exact same day Anthropic publicly announced the model.

How did they pull it off? The group was part of a Discord community dedicated to hunting down unreleased AI models. They used access held by an individual working at a third-party contractor hired by Anthropic. From there, they made an educated guess about the model's online location simply by analyzing the URL format Anthropic used for previous releases.

Since then, the group has been using Mythos regularly. To prove they had real access, they provided Bloomberg with screenshots and gave a live demonstration of the software in action. The source told the outlet that the group was only interested in playing around with new models, not wreaking havoc.

An Anthropic spokesperson told TechCrunch that the company is investigating the unauthorized access claim within the vendor environment, adding that there is no evidence yet of Anthropic's internal systems being compromised.

Why it matters

We think this leak exposes the most inconvenient truth in the frontier AI race: your model security is only as strong as your sloppiest vendor.

Anthropic has positioned itself as the gold standard of responsible AI safety. Mythos was supposedly so powerful that letting bad actors touch it was deemed a major risk. Yet a group of hobbyists slipped through the fence using basic credential sharing and URL guessing, bypassing the fortress before the dust even settled on the press release.

For enterprise developers and security leads, this is a loud wake-up call about AI supply chain risks. You can spend millions aligning a model and restricting its output, but none of that matters if your staging environments and third-party contractor credentials lack basic hygiene.

Anthropic got lucky this time because the intruders just wanted to mess with the software. But this incident makes it painfully clear that securing frontier AI is no longer just a model research problem, it is an everyday operational security nightmare.