Alleged Claude Mythos Breach Raises Questions About AI Security

InnovationEnterprise TechAlleged Claude Mythos Breach Raises Questions About AI SecurityByTim Keary,Contributor.Forbes contributors publish independent expert analyses and insights. Tim Keary is a reporter covering enterprise AI adoption.Follow AuthorApr 23, 2026, 10:58am EDT--:-- / --:--This voice experience is generated by AI. Learn more.This voice experience is generated by AI. Learn more.Hacker Claude Mythos (Photo by Jakub Porzycki/NurPhoto via Getty Images)NurPhoto via Getty ImagesAn unauthorized group of users gained access to Anthropic’s Claude Mythos model on the same day it was announced, according to a Bloomberg report released Tuesday. The users are said to be part of an online Discord group that searches for information about unreleased AI models.The report says that one of the users had privileged access as a worker at a third-party contractor to Anthropic. I requested comment from Anthropic on the incident, and a spokesperson responded via email that “we’re investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments.” The company also noted there was no evidence at this time that the reported activity extended beyond the third-party vendor environment or that Anthropic systems are affected. In any case, this alleged incident highlights the risk of frontier AI models being targeted by unauthorized actors. While Bloomberg notes the users don’t appear to have been malicious, if powerful models like Mythos were to fall into the hands of a cyber gang or nation-state, there could be serious security complications for enterprises. Frontier AI Companies Need Frontier Security The news comes just two weeks after Anthropic announced Claude Mythos Preview and Project Glasswing, a program that offered access to organizations involved in building or maintaining critical software. Greenwing made a big deal of controlling access to this powerful new model and the idea that a group of users could sid...