An 80% prompt injection bypass on Claude Code is today's urgent story.

From the editor

Two threads dominate today. First, agentic security is clearly not solved: an 80% bypass rate on Claude Code's auto mode is a five-alarm fire, arriving the same day Anthropic ships MHS to hook agents directly into biotech and quantum lab hardware. The blast radius of a compromised agent just got much larger. Second, the self-awareness research is quietly significant. Models that flag their own harmful outputs more accurately after misalignment suggests alignment has internal signatures we can measure. These threads will collide eventually.