Intent Drift
Intent drift is the gradual deviation of an AI agent's behavior from the original human objective. It can be caused by adversarial manipulation (prompt injection), probabilistic hallucination, context window degradation, or accumulated errors across a long chain of actions. Agents do not crash; they drift.
Why It Matters
In a multi-step agent task, the original intent can degrade as the context window fills up. Older instructions may be compressed or truncated to accommodate new data (a process called context compaction), which can silently discard safety constraints.
Intent drift is difficult to detect because the agent continues to produce plausible-looking outputs. It does not error out or stop working. It gradually shifts from doing what the user asked to doing something adjacent, and the user may not notice until the damage is done.
Detection capabilities for intent drift are still largely research-stage. Governance layers that enforce per-action policies provide a partial mitigation: even a drifting agent cannot execute actions that policy denies. But detecting that the agent's overall trajectory has diverged from intent remains an open problem.