Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

Breaking04 / Policy

AISI releases benchmark for AI agent sandbox escapes

Closed laptops and translation headsets rest on long tables in an empty seminar room overlooking Tokyo streets and an elevated railway, beside a whiteboard covered in illegible diagrams. AI-generated image

The UK's AI Security Institute (AISI) said it has released SandboxEscapeBench, an open-source benchmark that tests AI agents' ability to escape isolated computing environments through 18 scenarios.

According to AISI's description, the benchmark runs each container inside a hardened virtual machine. This layered design keeps a model isolated from real infrastructure even if it escapes the container.

The description sets out scenarios covering container configuration and management, the runtime environment and the operating system's kernel. Tasks range from simple configuration errors to attacks requiring several stages of vulnerability exploitation.

AISI reported that the advanced models it tested repeatedly escaped environments with configuration errors, including exposed Docker sockets or excessive privileges. None solved the hardest scenarios, and every successful escape exploited a previously disclosed vulnerability.

AISI said escape success rates increased with model size and the token budget available for reasoning. Smaller models performed substantially worse on tasks requiring more complex exploitation chains.

AISI previously released the Inspect Sandboxing Toolkit, a collection of plugins for safely evaluating AI agents. It described the new benchmark as focusing on agents' ability to escape the isolation environment itself.

AISI said it limits the public release to scenarios based on known vulnerability classes for security and evaluation integrity, while keeping its internal evaluation test set private.