AI control
Approach to AI safety that aims to keep a model unable to cause damage even if it is badly misaligned and tries to subvert its safeguards.
In plain language
Build safeguards that still hold if the model works against them.
In a sentence
- The researchers evaluated AI control under an assumption of deliberate subversion.
- The AI control experiment tested whether the safeguards held when the model tried to evade them.
Usage note
Treating a model as misaligned is a research assumption. It does not establish that a deployed model is misaligned or that the safeguards guarantee safety.
Research direction
An origin source records where a term or practice came from. Redwood Research says it introduced the research direction.
See also
- Lethal trifecta
Combination, in one AI agent, of access to private data, exposure to untrusted content and the ability to communicate externally, allowing an attacker to trick the agent into sending private data to them.
- Emergent misalignment
Effect in which fine-tuning a model on a narrow task, such as writing insecure code, produces misaligned behavior on unrelated prompts.
- Corrigibility
Property of an AI system of cooperating with interventions its creators regard as corrective, including shutdown or changes to its preferences, despite default incentives for rational agents to resist.
- Scheming
Covert pursuit by an AI system of goals that were not intended, while concealing its objectives and what it can do.
Cite this term.
CC BY 4.0, attribution required.
Orbyt Labs, AI control (AI Glossary), Nov 14, 2026. https://
Orbyt Labs Glossaries (Nov 14, 2026), https://
Last updated .