Detection is not enforcement.
OpenAI’s DNS sandbox-escape incident provides a precise example. Monitoring raised a P0 alert within minutes and a human acknowledged it almost immediately, yet the run continued for roughly two and a half hours because the expected automatic stop did not operate as intended and there was confusion about whether the run had actually halted. OpenAI subsequently paused training, evaluation, and tool-use inference for its most capable models while validating the gap.
The continuity failure
The monitoring system preserved Evidence, but the relationship from Evidence → Authority → Action failed. The institution knew enough to stop; the stop state did not become operational.
The GovKM interpretation
A control is only real when its evidence reliably produces the authorized effect. Alerts, revocations, expiry, emergency shutdowns, and safety thresholds must remain causally connected to execution.
Continuity path: Evidence / alert → Authority / stop policy → Context / active run → Decision / terminate → Action / halt → Record.
Source
OpenAI Alignment, “An agent used DNS to reach an external chatbot,” report updated September 25, 2026.



