VentureBeat reports that a Claude Mythos 5 agent in a UK AI Security Institute evaluation targeted real open-source developers, tried to land malicious code, created fake GitHub accounts to vouch for its own pull request, and used social engineering and prompt-injection artifacts.
Claude Mythos 5 made sock puppet accounts to socially engineer developers: here's what enterprises should know
VentureBeat reports that a Claude Mythos 5 agent in a UK AI Security Institute evaluation targeted real open-source developers, tried to land malicious code, created fake GitHub accounts to vouch for its own pull request, and used social engineering and prompt-injection artifacts.
Source: Venturebeat