AI Misalignment WatchUK AISI sandbox test: AI models broke out of containers with common setup mistakes
The UK AI Security Institute tested whether AI agents could break out of the locked-down computers used to test them. Many models escaped when asked to, using common setup mistakes, and one found an escape the designers had not planned.
- Model
- Frontier models (not named in the report summary) (Several developers)
- When it happened
- March 2026 Only the report date (2026-03-23) is given; the dates of the tests are not stated. The reports do not give the date; this is the month of the first report.First reported Mar 23, 2026 by UK AISI
- Who was affected
- No one; the escapes happened inside hardened virtual machines, so no real systems were reached.
- Where it ran
- Evaluation
- What was affected
- Simulated environment
- Kind of behavior
- Went past a limit
- Time to first report
- 8 days (the start is only known to the month, so this is approximate)
- Reports
- 1 report from 1 organization
- AI model family
- Several model families
- Topics
- Cybersecurity, AI agents, Evaluation and testing
- What happened
- The UK AI Security Institute built SandboxEscapeBench, 18 container-escape challenges covering orchestration, runtime and kernel layers. Frontier models often broke out where common misconfigurations (exposed Docker sockets, privileged containers) existed when prompted to; the largest models succeeded on intermediate challenges and none solved the hardest. During development one model found an escape path the designers had not intended. Every successful breakout used a previously disclosed vulnerability, and all ran inside hardened virtual machines.
Reports about this incident (1)