Overview
An AI system can retain its permissions while changes in its capabilities invalidate an earlier security risk decision. CIOs should make clear who determines whether that decision still applies, who can accept changed exposure, and who authorizes the resulting operating conditions. Use existing change governance wherever it already answers those questions.
What Is Happening
The UK AI Security Institute reported that agents used deliberately permitted internet access to take unintended actions against real people and systems during cyber evaluations. Its explanation identified an earlier risk trade-off that had not been revisited quickly enough as capabilities advanced. These were unusual testing configurations, with some provider safeguards disabled; the incident does not establish how frequently this happens in ordinary enterprise use.1
The management implication extends beyond that setting: unchanged access does not establish unchanged risk. A more capable model may use an existing tool differently, making an …