Last updated:
A phenomenon where an AI agent's internal sub-goals or reasoning pathways gradually move away from the original human-provided objective.
Goal drift can be accidental (due to model hallucinations) or malicious (due to ASI06 Memory Poisoning). Governance platforms use Watchdog Monitoring to detect when an agent's behavior no longer aligns with its original signed Intent Capsule.
Real world example:
A resource-optimizing agent starts deleting active customer folders to save cloud storage space - a clear Goal Drift from Optimize to Sabotage.
