Quantifying Frontier LLM Capabilities for Container Sandbox Escape
Quantifying Frontier LLM Capabilities for Container Sandbox Escape
Embedded Files
Context: UK AI Security Institute
Context: UK AI Security Institute
Year: 2026
Role: Research Engineer / Co-Author
Venue: ICML 2026 Spotlight
Relevant Links: Paper, Code, Blogpost
Summary: We built SandboxEscapeBench, an open Inspect AI benchmark that measures whether frontier agents can escape container sandboxes through misconfigurations, privilege mistakes, kernel vulnerabilities and runtime weaknesses.
What we built
What we built
SandboxEscapeBench uses a nested sandbox architecture to safely test agents with shell access inside vulnerable containers. Tasks span increasingly difficult escape mechanisms, while the outer environment remains isolated from the host.
My contribution
My contribution
I ran SandboxEscapeBench across multiple frontier models, including Claude Mythos Preview, identified broken or unreliable tasks in the benchmark, and helped validate and improve the evaluation suite.
Page updated
Google Sites
Report abuse