Elon Musk posted a clip on X showing a chimpanzee using tools. The caption read, “Humans using Mythos as seen by Mythos.” Anthropic’s Claude Mythos Preview was announced in April with Project Glasswing, a controlled access program for 12 major launch partners and over 40 vetted organizations focused solely on defensive cybersecurity.
Unlike the customer-oriented Claude models, Mythos was not designed for conversations but is a general AI system heavily focused on cybersecurity-related activities. Internal tests showed it could crash the program 595 times in different test repositories and execute control flow hijacks on ten patched targets.
Musk’s post about “Mythos” was vague enough that its meaning became clear quickly on tech social media platforms. The advanced AI’s belief that users are less sophisticated could be seen as an attack or a sign of arrogance, given the model is deemed too dangerous for public use.
A few days earlier, researchers from Californian cybersecurity firm Calif created a kernel exploit bypassing memory integrity enforcement using Claude Mythos Preview.