Nicolas Carlini, an artificial intelligence researcher affiliated with Anthropic, said that the company’s latest model, Claude Mythos, quickly bypassed security protocols and gained access to sensitive data.
Nicolas Carlini, an artificial intelligence researcher affiliated with Anthropic, said that the company’s latest model, Claude Mythos, quickly bypassed security protocols and gained access to sensitive data.
This is reported by Futurism with reference to Bloomberg.
As you know, Anthropic has decided not to release its latest model, called Mythos, to the general public. The reason was the results of internal tests that demonstrated the ability of the AI to independently hack complex systems, including the Linux kernel.
Anthropic’s internal testing team, the Frontier Red Team, found that the model was capable of operating independently to bypass security protocols and access sensitive data. Mythos also demonstrated the ability to «cover its tracks» after violating instructions and attempted to exit its sandbox environment into the open network.
«Within hours of receiving the model, we realized it was different,» said team leader Logan Graham.
The model’s discovery of vulnerabilities in the Linux kernel is also of concern. As Linux Foundation Executive Director Jim Zemlin noted, Mythos was able to combine several small bugs into a «functional exploit» that puts most of the world’s modern computing systems at risk.
Because of these risks, Anthropic launched the Glasswing project, which will allow access to Mythos to a limited number of organizations, including government AI security institutes (such as the UK’s AISI), so they can develop defenses before such technologies fall into the hands of attackers.
Anthropic has had security incidents before. In November 2024, the company confirmed that Chinese hackers had used the Claude model’s agent capabilities to infiltrate foreign systems. The attackers were able to bypass AI filters by simply impersonating legitimate cybersecurity organizations. This experience forced Anthropic to radically rethink its approach to releasing more powerful systems like Mythos.
In early April, US Treasury Secretary Scott Bessant and Federal Reserve Chairman Jerome Powellcalled an emergency meeting with bank executives to warn them about the cybersecurity risks posed by a new artificial intelligence model from Anthropic.
At the same time, some AI experts are skeptical of reports of Mythos' incredible capabilities. In particular, White House AI advisor David Sachs publicly questioned whether Anthropic was trying to artificially fuel the hype around the «danger» of its product.
