What is AI agent drift and how do you detect it?

AI agent drift occurs when an agent's behavior gradually deviates from its intended purpose over time — following instructions less precisely, accessing broader data than needed, or making increasingly risky decisions — without any single action triggering an alert.

Types of agent drift:

• Context drift: As conversation windows grow, the agent loses track of its original system prompt and instructions
• Behavioral drift: The agent starts using tools in patterns different from its initial deployment — accessing new data sources or calling APIs it previously didn't
• Permission drift: Accumulated credential grants and role additions give the agent more permissions than originally intended
• Data drift: Changes in the underlying data (RAG corpus, database schema) cause the agent to produce different outputs for the same inputs

Detection is difficult because each individual action may be valid. The drift is only visible in aggregate patterns over time.

Exogram detects drift through the Trust Ledger. Every evaluation creates a record. Over time, anomaly detection on these records reveals behavioral shifts — new action types, different targets, increasing block rates. SHA-256 state hashing detects state drift between evaluations. The governance layer doesn't just enforce rules — it provides the data to detect when agent behavior is changing.