The group took advantage of permissions from a third-party contractor to access Mythos, but did not use the model for cybersecurity research.
Unauthorized individuals accessed Claude Mythos, Anthropic’s new large language model, according to Bloomberg. Announced in early March, the AI had been made available to a restricted group of partners, but users from an online forum also managed to use the tool. The startup is investigating the incident.
According to a person familiar with the matter and Anthropic documents, several users from a private forum gained access to Mythos on the day the model was launched. The group continued using the AI regularly, but not for cybersecurity research, according to the company.
The group is part of a Discord channel dedicated to uncovering information about unreleased models. Based on Anthropic’s strategy in previous launches, they estimated the model’s online location. Additional details about the method were revealed in a recent data breach involving Mercor, an AI training startup that works with major developers in the sector.
The launch of Claude Mythos was marked by tension and caution from Anthropic. The company claims the model is capable of identifying and exploiting vulnerabilities in most operating systems and browsers — which is why access was restricted to a group of companies participating in the Project Glasswing program.
Until now, there had been no records of unauthorized access to the model. The leak reported by Bloomberg highlights the difficulty Anthropic faces in keeping the project secure and raises concerns that others may also be experimenting with the model without authorization.
How did they access Mythos?
To use Mythos, forum users combined different strategies. One method involved exploiting the access of an employee from a third-party company hired by Anthropic to evaluate model performance, along with the use of publicly available online investigation tools.
Anthropic investigates unauthorized access
“We are investigating a report of unauthorized access to Claude Mythos Preview through the environment of one of our third-party vendors,” said an Anthropic spokesperson in a statement. The company stated there is no evidence that the access extended beyond the vendor’s environment or affected Anthropic’s internal systems.
According to the startup, the group used the model in an exploratory manner without causing damage. There are no records of cybersecurity-related prompts being executed through Mythos.
