Unauthorized Group Accesses Anthropic's Mythos AI Model
A small Discord community has obtained access to Anthropic's Claude Mythos Preview without permission. Members of this group have used the model for two weeks. Bloomberg first reported the incident based on details from an unnamed participant.
Details of the Unauthorized Access
The group includes a third-party contractor who works for Anthropic. This individual shared access details with forum members. They combined the contractor's credentials with basic online research methods to reach the model. Access began on April 7. That date matches Anthropic's announcement of Mythos availability to select partners for testing.
Group members provided Bloomberg with screenshots and a live demo as proof. They have tested Mythos regularly but avoided cybersecurity-related tasks to stay under the radar. The publication also notes that this same group reached other unreleased Anthropic models. The Discord channel focuses on tracking unpublished AI systems.
They located Mythos by drawing on details from a recent Mercor data breach. That event exposed information about Anthropic's model structures. Group members made a calculated prediction about Mythos's web address based on those patterns.
Capabilities of the Claude Mythos Preview
Anthropic describes Claude Mythos Preview as a general-purpose AI tool built for cybersecurity. It can detect and take advantage of weaknesses in all leading operating systems and web browsers. Users must direct it to perform these actions.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
The company restricts official use to a few organizations under Project Glasswing. Participants include Nvidia, Google, Amazon Web Services, Apple, and Microsoft. Government agencies have shown interest in the technology. Anthropic has decided against a public launch. Officials worry that bad actors could turn the model into a weapon.
Anthropic started in 2021. Former OpenAI leaders founded the firm to prioritize AI safety. Its Claude series of models emphasizes alignment with human values. Mythos fits into efforts to apply AI for defensive security purposes. Recent reports highlight tensions, such as the Pentagon viewing Anthropic as a supply-chain risk. This model might help repair those relations.
Anthropic's Official Statement
An Anthropic representative commented to Bloomberg. "We're investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments." The company sees no signs that the breach affects its core systems. It appears confined to the vendor's setup.
Anthropic maintains tight controls on previews like Mythos. The firm routinely vets partners and monitors for risks. This event underscores challenges in securing advanced AI amid growing demand from tech giants and governments.
Broader Context on AI Security
Incidents like this raise questions about protecting cutting-edge models. Anthropic joins peers in limiting access to high-risk tools. The Mythos leak follows patterns seen in other breaches, such as Mercor. That earlier event leaked model format data, aiding unauthorized entry.
The Discord group's actions highlight how determined outsiders combine insider info with public sleuthing. Anthropic now faces pressure to strengthen vendor oversight. No evidence shows misuse yet, but the potential remains high given Mythos's power.