Anthropic is currently looking into allegations that a small group accessed its Claude Mythos model, a cybersecurity tool deemed too potent for public release. A statement from the company noted that the investigation was prompted by a report indicating unauthorized access via a third-party vendor's environment. This situation arose following a Bloomberg article revealing that users in a private forum had managed to interact with the model without proper permissions.
Concerns regarding the capabilities of Mythos have been raised, although the UK's National Cyber Security Centre has suggested that advanced AI tools could offer significant benefits if safeguarded properly. There is no current evidence indicating that malicious entities have exploited the model, and Anthropic maintains its systems have not been compromised.
According to Raluca Saceanu, CEO of SmartTech, the access was likely due to misuse of permissions rather than a traditional hacking incident. The individual involved reportedly had legitimate access through a prior engagement with a contractor. Despite their access, it appears the group has not utilized the model for illicit purposes, aiming to avoid detection.