Semantic Firewall

An emerging name for a control layer that evaluates the meaning and context of inputs, outputs, or tool requests before allowing them to proceed. It can look for prompt injection, attempted exfiltration, policy conflicts, malicious instructions in retrieved documents, or inappropriate use of a tool.

The term does not describe one market-standard pattern and does not create a reliable security boundary on its own. Semantic classifiers are probabilistic and can be bypassed. Effective protection requires defense in depth: keep untrusted data separate from instructions, use least-privilege permissions, deterministically validate arguments, require human confirmation when risk demands it, and collect telemetry on blocks.