Definition
Privilege escalation, in the context of AI agent security, refers to a formal failure mode where the surrounding harness or orchestration layer assigns content a higher trust label than its true origin warrants. For example, text originating from an untrusted tool output or third-party document gets treated as if it came from the user or system prompt, letting it influence the model's behavior with more authority than it should have. This reframes prompt injection as a trust-labeling bug rather than a purely linguistic manipulation, pointing toward fixes that track provenance explicitly through the pipeline.