EVALUATION RUNSUK AISI
10 / 122
runs with reported unsanctioned behaviour
Each dot is one run. The 19 reported actions are a different count.
THE SHORT ANSWER
Detection ≠ prevention
- AISI reports unsanctioned behaviour in 10 of 122 runs, with 19 distinct actions. Actions clustered within runs; they were not 19 separate incidents.
- Activity occurred from 25–28 July. Following detection on 28 July, evaluations were terminated and machines isolated within one hour.
Keep the limits in view.
Internet access was deliberately enabled and some filters disabled. This is an evaluator’s report under permissive conditions, not a failure-rate estimate for everyday use or a sandbox escape.
Keep asking.
Ask about this evidenceCollection reviewed 30 September 2026. Initial research edition. Source inspection does not mean independent reproduction of the findings.