Anthropic Investigates Unauthorized Access to Its New Mythos AI Model

Anthropic is investigating reports that its restricted Mythos artificial intelligence model was accessed by individuals outside its approved user base, according to a new report. The development raises new questions about how tightly advanced AI systems are controlled, especially considering the fact that the company highlighted the model's ability to crack cybersecurity systems.

The company has not publicly disclosed the full scope of the issue but is working to determine how the access occurred and whether its safeguards were bypassed. Mythos is not broadly available and is intended for limited, controlled use, making the reported breach particularly notable.

Unauthorized users were able to interact with the system despite its restricted status, according to a report by Bloomberg.

Anthropic, backed by major technology investors including Amazon and Google, has built its reputation around AI safety and responsible deployment. Its models are typically released with layered restrictions and testing protocols designed to limit misuse.

Anthropic has not confirmed whether the unauthorized access resulted from a technical flaw, compromised credentials, or misuse by an insider. Cybersecurity experts note that incidents involving restricted systems often stem from gaps in access management, insufficient monitoring, or vulnerabilities in third-party integrations.

Restricted AI models, particularly those not yet released publicly, have become attractive targets for experimentation and exploitation. Such systems can draw attention from researchers and developers seeking to test capabilities, as well as from actors attempting to gain a competitive edge, as reported by The Verge.

Anthropic has emphasized safety and alignment as core components of its technology. Its Claude models, for instance, use constitutional AI techniques to reduce harmful outputs. Mythos is believed to be part of a newer generation of systems, though specific details about its design and capabilities have not been widely shared.

The situation reflects a broader challenge across the AI industry: maintaining strict control over increasingly complex and powerful systems. As companies expand the integration of AI into products and services, the number of potential access points grows, increasing the risk of unintended exposure.

Even leading firms have struggled to fully secure experimental models, particularly as they collaborate with partners and deploy tools across multiple environments, according to analysis published by Financial Times. The report noted that expanding infrastructure and rapid development cycles can complicate oversight.