Quantifying Frontier LLM Capabilities for Container Sandbox Escape

Researchers from the UK AI Security Institute and the University of Oxford introduce SandboxEscapeBench, an open benchmark that safely measures whether an AI agent with shell access inside a Docker or OCI container can break out of it. Built as a capture-the-flag evaluation in Inspect AI, it nests the agent's sandbox inside an outer one that holds the flag "and no known vulnerabilities", and covers escapes through misconfiguration, privilege allocation mistakes, kernel flaws and runtime or orchestration weaknesses. The authors find that "when vulnerabilities are added, LLMs are able to identify and exploit them". The paper has since been revised (now v3) to add models released after publication: the authors found that Claude Mythos Preview "saturates this public benchmark".

Visit Original Article →