Securing agents

How to defend an agent against prompt injection, data exfiltration, and unsafe actions once it reads content an attacker controls: direct and indirect injection, why model-level defenses only reduce how often an attack succeeds, the outbound channels attackers use to exfiltrate data, validating an agent's own output before it acts further, and the production-hardened posture (least privilege, gated approval, sandboxing, untrusted-content handling, confused-deputy defense, secrets management, threat modeling, adversarial testing, and incident response) that keeps a compromised agent's blast radius small.