According to the report, the users were members of a private Discord channel and managed to access the model on the same day it was announced. They allegedly used access credentials connected to one member who worked as a contractor for Anthropic, along with publicly available information that had surfaced in a data leak involving the AI startup Mercor.

Bloomberg says the group did not use Mythos for cyberattacks. Instead, they reportedly used it for harmless tasks, such as building simple websites.

The same source also claims that the group may have gained access to several other unreleased Anthropic models. Anthropic said it is investigating the incident. So far, the company says there is no indication that the access went beyond the environment of the outside service provider or that Anthropic’s internal systems were affected.