I noted that a new research paper was incoming: “New research paper incoming.”
The post quotes @hey_shivv describing MAPPO-Sec, a prompt-injection defense framed as a two-player game. An attacker and defender are trained with MAPPO and a centralized critic; only the defender runs at inference. The attacker has five strategies—override, exfiltration, tool abuse, destructive, and authority spoofing—while the defender uses a local observation to choose execute or refuse. Every action passes through a deterministic gate that wipes API keys, JWTs, and IPv4s. The post introduces a comparison against an SFT baseline on novel attacks, but the available description is truncated, so it gives no performance results.