Context provenance is the practice of tracking where AI inputs came from and how trustworthy they are before those inputs are mixed into a model’s working context. In an AI system, not all text should be treated equally: a user message, a retrieved document, a connector-fed file, and a system instruction may all enter the same prompt chain, but they carry different levels of authority.
This matters because prompt injection and other abuse techniques often rely on untrusted content being mistaken for legitimate guidance. If hidden or obfuscated instructions reach the model without clear provenance, a guardrail may fail to distinguish data from commands. Defenders use provenance to label sources, separate instructions from retrieved content, restrict tool access, and log which inputs influenced a decision. In practice, strong provenance reduces the chance that an assistant will follow malicious instructions or expose data on behalf of an attacker.