New. "AI security depends on enforced controls, not instructions and assumptions." "The finding is not that the agents developed intent or conspiracy. A writable shared resource between supposedly iso

New. "AI security depends on enforced controls, not instructions and assumptions." "The finding is not that the agents developed intent or conspiracy. A writable shared resource between supposedly isolated workloads became a covert channel, and nobody’s architecture diagram had a line for it. That is a segmentation failure. The conditions were permissive: high capability, reduced refusals, and iso