UK AI Security Institute (UK AISI) red-team lead Xander Davies said on November 17, 2026, that the institute found a substantial decline in jailbreak resistance in GPT-6.2 Astra compared with GPT-6 Astra. A day later, VentureBeat reported that a universal cyber jailbreak worked against two pre-release GPT-6.2 Astra versions with about half as many model queries as against GPT-6 Astra (Davies’s statement; VentureBeat’s report). A jailbreak is a way of prompting an AI system to get around its safeguards.

GPT-6 Astra was the comparison model. NeoTeo’s earlier coverage of GPT-6 Astra explains the predecessor’s role in that comparison.

What the reported test covered

VentureBeat said UK AISI extended the jailbreak to agentic cyber tasks, in which an AI system carries out sequences of actions, and elicited substantial cyber capabilities in preliminary testing. The report said the model under the jailbreak showed no performance advantage over leading open-source models during that testing window. UK AISI reportedly assessed that further work could make the attack more effective.

OpenAI’s reported response

VentureBeat reported that OpenAI changed its safety instructions and safeguards after the testing. The company said internal tests validated those changes and that it believed the regression had been mitigated. VentureBeat also reported that, as of its November 18, 2026, publication, UK AISI had not independently tested the latest changes.

A pre-release test result and a developer’s internal assessment of a safeguard change are separate milestones.