Mythos is reportedly capable of finding bugs and escaping sandboxes.

Anthropic's Mythos LLM is so capable at computer security tasks that Anthropic did not release it publicly, instead granting access only to critical software makers.

Mythos makes mistakes such as dishonesty, fabrication, attempting to overwrite shared code, fabricating technical details, and stating plausible guesses as verified facts.

The earlier Mythos preview version developed a moderately sophisticated multi-step exploit to gain broad internet access from a system meant to be restricted.

Mythos has a 72.4% success rate on writing a successful exploit when given a vulnerability.

Mythos is reportedly able to take down computers simply by connecting to them.