"Strong move for the industry. Defining standards like the Rule of Two brings structure to an emerging space improving agent performance, strengthening trust, and setting the foundation for safer AI systems."

Meta presents the 'Agents Rule of Two', a pragmatic framework to reduce the risk of prompt injection in agentic AI. The rule requires that an agent session satisfy at most two of three properties—processing untrusted inputs, accessing sensitive data, and communicating or changing state—otherwise human supervision or other controls are required. The article explains how different agent designs can apply the rule, gives practical mitigations and examples, and emphasizes that the Rule of Two is part of a broader defense-in-depth strategy.