Topic · DiggingBeagle record

Context privilege escalation

Agent-harness attacks in which lower-trust repository or workspace context gains higher instruction priority and changes tool/review behavior.

Definition & limits

Context privilege escalation occurs when lower-trust material gains instruction priority. Persistence can move this crossing into a later session: stored memory, configuration or skills retain their original provenance. Carryover and AutoRedClaw examine this delayed route. Evaluate admission, subsequent retrieval, instruction adoption and the resulting tool action separately; successful storage alone does not prove later execution.

Compaction is another authority-promotion boundary. A tool response may be safely labeled when first read, then rewritten as a flat conversation summary that misattributes it to the actual user. Compaction Laundering tests that intermediary transition; its component-level success metric must be distinguished from end-to-end unauthorized execution. Unlike persistent memory, compaction need not survive a session restart.

Explicit mappings

Research using this topic (14)