Safety
157d agoSession Risk Memory Enhances Agent Safety
Session Risk Memory (SRM) augments deterministic safety gates by tracking agent trajectories, detecting distributed attacks that split harmful intent across compliant steps. By maintaining a semantic centroid and accumulating risk via exponential moving averages, SRM offers lightweight, trajectory‑level authorization.
The Signal
This approach strengthens AI safety worldwide, enabling secure deployment of autonomous agents across industries and reducing systemic vulnerabilities.