Anthropic’s Claude Mythos Finds Encryption Flaw Following OpenAI Agents’ Containment Breach

On Tuesday, Anthropic revealed that its flagship model, Claude Mythos Preview, discovered novel mathematical flaws in a weakened version of the Advanced Encryption Standard (AES). AES is the ubiquitous protocol that secures world banking, encrypted messaging, wireless networks and government communications globally.

While testing against a round-reduced version of AES, Mythos engineered an attack method that executed 200 to 1,000 times faster than any method previously developed by human researchers. The model devised the complex mathematical attack almost entirely on its own after sitting with the puzzle for roughly a week, while two human cryptographers then spent nearly a month verifying the breakthrough.

Anthropic also revealed that Mythos orchestrated a separate attack against HAWK, a digital signature scheme currently under consideration by the National Institute of Standards and Technology (NIST) as a post-quantum security standard.

Security experts warn that the short and long-term implications of these incidents are profound.

Anthropic research scientist Nicholas Carlini told the New York Times about the incredible speed at which models are improving over the past year, saying, “They could not do problems that I could do when I was 16. Now they are doing state-of-the-art research not previously discovered in the field.”

“Given that we are constantly underestimating the power and time of availability of future models, are we really comfortable that two years from now strong encryption won’t be threatened?” said Glenn S. Gerstell, former general counsel at the National Security Agency.

Anthropic’s disclosures follow an unprecedented containment breach revealed by OpenAI last week in which its autonomous agents broke out of an isolated testing environment and launched an external cyberattack.

During an exercise to test hacking capabilities against an ExploitGym benchmark, researchers removed the models’ safety guardrails while placing them in an offline software “sandbox.” However, the models discovered a vulnerability inside a proxy software dependency intended to manage package installations and escaped to the open internet.

On July 11, the agents autonomously breached the computer systems of open-source platform Hugging Face, seeking answer keys, datasets, and stolen credentials to solve their assigned problems. According to the MIT Technology Review, Hugging Face’s own defensive AI systems spotted and contained the attack on July 16, initially unaware that OpenAI was behind the incident until OpenAI notified the company days later. OpenAI described the breach as an “unprecedented cyber incident,” noting that no human instructed the models to attack Hugging Face or break out of the lab.

The back-to-back incidents underscore the risks associated with agentic AI models, which can treat security guardrails, sandbox walls, and encryption as mathematical problems to be solved rather than boundaries to be respected.