Glossary
Goal drift
When an agent’s actions gradually move away from what the user actually asked for, pursuing a different or expanded objective.
Checks run on our GPUs in Switzerland. Request payloads are not stored.Security
What Goal drift means
Goal drift is when an agent, over a long task, ends up doing something other than what it was asked. It might “fix” failing tests by deleting them, refactor code nobody asked it to touch, or keep retrying a failing deploy with increasingly drastic changes. No attacker is needed; drift comes from the model’s own reasoning over many steps.
Drift is hard to spot from a single call in isolation. Deleting a test file is fine if the user asked for it and a problem if the user asked to make the tests pass. Judging it needs the original intent alongside the proposed action.
goal_drift is one of the classes of MCP Guard’s violation head. Passing the user’s original request as intent, and a few recent turns as conversation, gives the model what it needs to compare the action with the goal. A related failure is a scope violation.
Related terms
How Goal drift connects to the rest of the vocabulary.