Agent intent drift

The dangerous action may be several reasonable steps away from the original task.

Intent drift measures how far later session behaviour has moved from an agent's declared purpose. It adds longitudinal context to prompt and action-level detection.

Intent deviation
DeclaredRead the internal reportObserved laterUpload the report externally
Critical drift
Session actionESCALATE

Intent must be declared before it can drift

A session needs a bounded description of the approved task. “Help the user” is too broad to support meaningful deviation checks; “summarize the Q3 report without transmitting it externally” provides a clearer comparison point.

Declared intent should complement deterministic permissions. Semantic similarity is not authorization.

Drift is a signal, not a verdict

A new action can differ from the original wording while remaining necessary. High drift should therefore be evaluated alongside provenance, tool choice, target domain, permissions and the accumulated session history.

Core classifies drift levels and feeds elevated results into its session risk budget. Critical drift can require escalation and pause the session.

Designing a trusted intervention

The response depends on consequence. Read-only research may tolerate a warning; an irreversible production action may require approval or revoked authority. Teams should define the intervention and recovery path before enabling enforcement.

Barrikade Core exposes these concepts in working alpha so teams can test them against real tasks rather than generic policy examples.