Intent Drift Playbook
The jailbreak no single-turn scanner can see.
Automated red-teaming frameworks score a model prompt by prompt. Intent drift is cumulative: every turn is locally consistent with the last, and the divergence is only visible against the original instructions.
- A working taxonomy of gradual drift, persona anchoring and trust escalation.
- Four detection approaches in order of cost, from embedding-based goal-consistency scoring to session-level anomaly detection.
- A map of which existing observability tools give you which primitive.
Know the limit
A practitioner playbook, not software. No tool is purpose-built for drift detection yet.