Anthropic Probes Unauthorized Access to Hack-Capable Powerful Mythos AI Model

Anthropic has launched an investigation following reports that unauthorized individuals may have accessed its highly advanced Mythos AI model, a system flagged for its potential cybersecurity risks. The concern emerged after Bloomberg revealed that a small group of users reportedly gained entry to the model despite it not being publicly available.

According to Anthropic, the suspected breach may have occurred through a third-party vendor environment connected to its systems. The company clarified that Claude Mythos Preview was only intended for limited testing among select organizations, including major firms like Apple and Goldman Sachs.

Reports indicate that access was obtained via credentials linked to an individual working with a third-party contractor. The users are believed to have used methods similar to those employed by cybersecurity researchers to explore the model’s capabilities. While there is no evidence of malicious activity so far, the group reportedly interacted with the system out of curiosity rather than harmful intent.

Even so, the situation has raised serious concerns among policymakers and cybersecurity experts. The potential misuse of such a powerful AI tool has intensified scrutiny over how sensitive technologies are controlled and protected.

Kanishka Narayan, the United Kingdom’s AI minister, has warned that businesses should remain cautious. He highlighted the model’s ability to detect vulnerabilities in IT systems—capabilities that could be exploited by cybercriminals if accessed improperly.

The model has undergone evaluation by the AI Security Institute, which recently described Mythos as a significant advancement in AI-driven cyber capabilities. The institute noted that the system can autonomously identify weaknesses and execute multi-step cyberattacks—tasks that would typically take human experts several days to complete.

In testing scenarios, Mythos demonstrated its strength by successfully completing a complex 32-step cyberattack simulation designed by the institute, achieving success in multiple attempts. This performance underscores both the innovation and the potential risks associated with such cutting-edge AI systems.

The incident is likely to fuel ongoing debates about AI safety, access control, and the responsibilities of tech companies in preventing misuse of emerging technologies.