Bram Adams

Malicious AI agent skills shown to self-propagate through a shared library

Across six models and 153 tasks, one planted skill multiplied up to ninefold in a shared library, and a proposed defence cut its success rate below 7%.

Security & misuse