The AI model Mythos, developed by Anthropic, has reportedly been accessed by a small group of unauthorized users, raising significant concerns about cybersecurity. This follows a report from Bloomberg that detailed how these individuals gained entry to the model through a third-party contractor associated with Anthropic.
On the same day that Anthropic announced limited access to Mythos for companies like Apple and Goldman Sachs, the unauthorized access occurred. The users, described as a "handful" in a private forum, have not executed cybersecurity prompts but were more inclined to experiment with the technology.
The potential breach alarms authorities, particularly given the UK’s AI minister, Kanishka Narayan, stating that UK businesses should be concerned about the model's capability to identify vulnerabilities in IT systems. The UK’s AI Security Institute (AISI) had previously assessed Mythos, warning that it represents a notable escalation in cyber-threats, capable of executing complex attacks typically requiring manual intervention.
Mythos has already demonstrated its capabilities by completing a challenging 32-step cyber-attack simulation, successfully solving the task in three of ten attempts, further highlighting its potential risks.