Elon Musk has taken aim at Anthropic's most powerful and most restricted AI model yet, posting a clip of a chimpanzee using tools on X alongside the caption that stated, "Humans using Mythos as seen by Mythos."

Claude Mythos Preview was announced in April alongside Project Glasswing, a controlled access programme that limits the model to 12 major launch partners, including AWS, Apple, Cisco, Google, and Microsoft, plus over 40 additional vetted organisations, all for defensive cybersecurity work only. Anthropic has stated it will not make the model publicly available.

Unlike Anthropic’s customer-orientated Claude series of AI models, Mythos was never designed for conversations. This is a general AI system, however, with an extraordinary emphasis on cybersecurity-related activities, particularly the detection and exploitation of zero-day software vulnerabilities on a scale that was unprecedented compared to all previous AI systems.

In the internal tests conducted by Anthropic, the AI managed to crash the program through entry points 595 times in different test repositories and managed to execute control flow hijacks on ten patched targets.

While Musk’s chimps did not mention Anthropic by name, the "Mythos" mention is sufficiently vague that everyone could tell what he means within seconds on tech social media platforms.

The fact that an AI this advanced believes that it needs to treat its users as less than sophisticated can have multiple meanings, either as an attack on the people who use the service or as the arrogance of creating something deemed too dangerous to be used.

Not long ago, he described Claude as being “misanthropic and evil" and ideologically biased and the company as hating the West because of a disagreement with the Pentagon about how they use military AI.

Only a few days ago, a team of researchers from Californian cybersecurity company Calif managed to create a kernel exploit bypassing memory integrity enforcement using Claude Mythos Preview.