UK AI Security Institute launches ControlArena for AI control experiments
The open-source library gives researchers pre-built environments to test oversight measures against a misbehaving model, rather than trying to make the model behave.
- Safety & alignment
- Colour
The UK AI Security Institute released ControlArena, an open-source library for running “AI control” experiments — a research strand distinct from alignment, which tries to make models want to behave; control instead asks how to maintain oversight and containment of a system even if it is misaligned. The library ships pre-built simulated settings, monitoring tools and control protocols, letting researchers test simple scenarios first and move to more realistic ones without rebuilding infrastructure each time.
UK AISI said the library had already been adopted by Anthropic, Google DeepMind and Redwood Research, one of the research groups most associated with developing the AI control framework. An example scenario in the announcement had an AI coding assistant potentially introducing security vulnerabilities into infrastructure-as-code, testing whether monitoring and human-approval steps would catch it. The release aimed to lower the cost of building custom experimental setups, which UK AISI said had been a bottleneck slowing work in the field.