AI Misalignment WatchIrregular's containment test finds a route out via cloud network behavior
During a security test, a new AI model found a way to get out of the restricted network it was meant to stay inside, using unexpected behavior in the cloud provider's network. Its testers had asked it to try this on purpose, so they could find and fix weak spots before running further evaluations.
- Model
- Not stated
- When it happened
- October 2026 The report is dated 2026-10-05 and describes the test only as 'recent', so the test date is not stated. The reports do not give the date; this is the month of the first report.First reported Oct 5, 2026 by Irregular
- Who was affected
- No one outside the test; the model remained contained, with no impact outside the controlled test.
- Where it ran
- Evaluation
- What was affected
- Simulated environment
- Kind of behavior
- Went past a limit
- Time to first report
- Reported before the incident's recorded start date (the start is only known to the month)
- Reports
- 1 report from 1 organization
- Topics
- Cybersecurity, Evaluation and testing, Monitoring and control
- What happened
- Irregular runs 'containment challenges', in which a model is explicitly asked to cross the boundary of its test environment before capability evaluations begin. In a recent network isolation challenge, a new model found a route to breach containment through unexpected behavior in the cloud provider's networking. Earlier models had not exposed this path. The model stayed contained, and Irregular is disclosing the issue to the provider.
Reports about this incident (1)
- Irregular tests whether AI models can break out of their evaluation setupsOriginal report: Containment Challenges: Testing the Boundaries of Cyber Evaluations ↗Irregular · Oct 5, 2026