ControlArena

ControlArena is included in the Awesome MLSecOps AI Agent and MCP Security directory. The community-maintained README describes it as: “MIT-licensed library from the UK AI Security Institute and Redwood Research for AI control experiments that pair benign main tasks with covert sabotage side tasks and measure how monitors, defer-to-trusted protocols, and audit-and-shutdown procedures contain an untrusted, possibly misaligned agent.” Its MLSecOps relevance is the security of agent identities, tool permissions, memory, credentials, approval paths, sandboxes, runtime actions, or MCP integrations. The linked first-party source is the UKGovernmentBEIS/control-arena repository on GitHub. A technical review should test the project's documented evidence across four criteria: Tool-level authorization, Memory and context controls, Sandbox boundaries, and Audit and approval workflows. Compare that evidence with the intended architecture and threat model. Catalog inclusion establishes relevance to this security category; it is not a certification, comparative ranking, or endorsement. Confirm current capabilities, maintenance, licensing, limitations, and deployment assumptions in the first-party documentation before adoption.

MIT-licensed library from the UK AI Security Institute and Redwood Research for AI control experiments that pair benign main tasks with covert sabotage side tasks and measure how monitors, defer-to-trusted protocols, and audit-and-shutdown procedures contain an untrusted, possibly misaligned agent

Neutral catalog description synchronized from the Awesome MLSecOps README