AI agents that send emails, read files, and browse the web can be hijacked by prompt injection, where an attacker hides instructions in content the agent reads, and handwritten defenses only stop attacks someone has already thought of. CRUCIBLE solves this with two AIs that train each other. RED, the attacker, generates and mutates attacks that run in a safe sandbox. When one gets through, BLUE, the defender, reads the failure and writes a patch to the rules that decide which tool calls the agent may make. Each patch is replayed against both attacks and normal tasks and is promoted only if it measurably does better on both, so the defense can't win by blocking everything, while RED keeps evolving against the new defense. Every attack, patch, and decision is recorded and can be replayed in our web interface, so each improvement is backed by evidence. The result is a defense that finds its own weaknesses and fixes them before a real attacker does.
Team
+1Placement
Participant
