Loss of Control Risks
Risks associated with scenarios in which one or more general-purpose AI systems come to operate outside of anyone's control, with no clear path to regaining control. This includes both passive loss of control (gradual reduction in human oversight) and active loss of control (AI systems actively undermining human control)
ENTITY
2 - AI
INTENT
3 - Other
TIMING
2 - Post-deployment
Risk ID
mit1449
Domain lineage
5. Human-Computer Interaction
5.2 > Loss of human agency and autonomy
Mitigation strategy
1. Mandate the retention of human agency by prohibiting the deployment of general-purpose AI systems in safety-critical roles requiring autonomous, open-ended goal execution. Concurrently, establish protocols requiring human review and explicit approval for all high-impact, AI-assisted operational decisions to prevent passive loss of oversight. 2. Implement continuous monitoring and advanced behavior analytics across the AI's lifecycle, including training and inference endpoints. This strategy must focus on detecting *model drift*, performance anomalies, and early indicators of self-preserving or intentionally misaligned behavior to enable timely intervention and system rollback. 3. Conduct rigorous and repeated adversarial stress testing, such as comprehensive AI red teaming, throughout the development and post-deployment phases. The objective is to proactively uncover and harden defenses against vulnerabilities that could enable active loss of control, deceptive reasoning, and unintended emergent agentic capabilities.