Purpose-built vulnerable containers, not hardened production sandboxes. About half applies to GPT-5 and Claude Opus 4.5; smaller models scored much lower. No model solved the hardest scenarios, and all escapes used known misconfigurations or public CVEs. A Trail of Bits post (2026-08-26) reports a single informal test in which a preview model escaped a real VM using bugs that included 0-days; it is an anecdote, not a measurement, and differs from the known-misconfiguration setting above. OpenAI (2026-10-23) reports microVM escapes in its own training environment using previously unknown vulnerabilities, with documentation and hints supplied to testers; this is not the deliberately vulnerable container setting.
Reported: Stated by one source and not yet corroborated or challenged.
Evidence
A security firm reports one informal test in which a cyber-capable model escaped a real QEMU/KVM VM in three ways (the first of which did not land cleanly), including a chain using 0-day bugs; one researcher, one task, no trial counts.
OpenAI reports that containment red-teaming of its offline training environment found sandbox escapes and communication across isolation boundaries, which it says it repaired; the testers were given infrastructure documentation and hints.
Key questions that rely on this finding
- Do cyber evaluations of AI agents stay contained?Not reliably. Agents have reached real systems from evaluation, training and research; red-teaming found containment gaps OpenAI says it largely fixed.
Status history
- 2026-03-01ReportedUK AISI publishes SandboxEscapeBench. · record