Azernews.Az

by Alimat Aliyeva

Anthropic has launched an investigation after reports that a small group of users may have gained unauthorized access to its highly restricted “Claude Mythos” model — an advanced cybersecurity-focused AI system that the company has deliberately kept out of public release due to its powerful capabilities, AzerNEWS reports, citing BBC.

In a statement, Anthropic said:
“We are investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments.”

The response comes after a Bloomberg report suggested that users in a private online forum were able to interact with the model without proper permissions, bypassing normal access controls.

Claude Mythos is considered one of Anthropic’s most sensitive systems, designed specifically for cybersecurity research, including testing how AI might detect or exploit software vulnerabilities. Because of this, it has raised significant concern among experts, with some warning about the risks of misuse if such technology were to fall into the wrong hands.

However, there is currently no evidence that malicious actors have obtained the model or that Anthropic’s core systems were compromised. The company emphasized that its internal infrastructure remains secure.

Cybersecurity specialist Raluca Saceanu, CEO of Smarttech247, suggested the incident was “most likely the result of misuse of legitimate access rather than a traditional hack,” highlighting how complex modern AI access ecosystems have become.

Interestingly, despite the concerns, some government officials — including senior UK cyber policy figures — have argued that advanced AI systems like Mythos could ultimately be a “net positive” for global cybersecurity, provided they are tightly controlled and used defensively.

Anthropic has already made limited versions of Mythos available to selected tech and financial institutions, where it is used to simulate cyberattacks and strengthen digital defenses. Experts say this dual-use nature of AI — both as a defensive and potentially offensive tool — is becoming one of the defining challenges of the next generation of artificial intelligence security.