In a very carefully worded story from Bloomberg, Anthropic confirmed Tuesday that it has received a report that an unauthorized mystery group is accessing the cloud mythos — a model it says is too dangerous to release. “We are investigating a report claiming unauthorized access to Cloud Mythos Preview through one of our third-party vendor environments,” an Anthropic spokesperson statement to Bloomberg said.
Bloomberg confirmed the apparent breach by viewing a live demo and screenshots sent by a member of the group responsible for the unauthorized access.
In apparently vague language, Bloomberg explains that an unnamed source says they are members of an unnamed group who abused their access “as workers of a third-party contractor for Anthropic” and used “Internet investigative tools commonly used by cybersecurity researchers” to gain some kind of access to the model.
But don’t worry, this secretive group that apparently has access to the world’s most dangerous technology is “interested in playing with new models, not in wreaking havoc with them,” the source bluntly explained to Bloomberg.
The sequence of events in a clear violation looks like this:
- A Discord group exists that uses bots on GitHub to mine information about unpublished AI models
- AI training startup Mercor had a data breach
- The group combined information from the Mercator breach with access to Bloomberg’s source because they work for an Anthropic contractor.
- This allowed the group to estimate the online location of the cloud mythos.
- The group has been openly messing with the cloud mythos since April 7, the day Project Glasswing was announced.
So to recap: Anthropic says it has the world’s scariest AI model, and for what it’s worth, a lot of powerful institutions believe it. If we take Anthropic at its word, then we are all trusting that we will not abuse this power that it and only it controls. However, some unknown organization has gained access to this scary AI model, but if we take Them According to their words, they have only used it for some vibe coding tests and they swear they are not doing anything bad with it.
<a href