
UK AI Safety Institute finds universal GPT-6.2 jailbreak in half the queries, OpenAI addresses regression
The U.K. AI Security Institute disclosed a universal jailbreak against pre-release GPT-6.2 Astra models, halving the queries needed compared with GPT-6 Astra, and OpenAI has addressed the regression. The disclosure matters because frontier models are being positioned for cybersecurity and code-patching roles, where safety failures can undermine defensive workflows.
OpenAI has addressed a jailbreak robustness regression in two pre-release versions of GPT-6.2 Astra after the U.K. AI Security Institute said it developed a universal jailbreak against the models in half the queries needed for GPT-6 Astra. The disclosure appeared in a weekly threat bulletin that also listed WordPress security fixes, Google's Gemini 4 Neon rollout and OpenAI's GPT-6.2 Astra and Sol models. The finding centers on a safety gap in a system OpenAI says meets its Critical cybersecurity threshold.
The technical claim is narrow but specific: the jailbreak worked against two pre-release GPT-6.2 Astra variants and required half as many queries as a comparable attack path for GPT-6 Astra. OpenAI's own announcement says GPT-6.2 Astra and Sol meet the Critical threshold under its Preparedness Framework for cybersecurity, citing improved vulnerability discovery, patching and token efficiency. That combination makes the AISI disclosure more consequential than a routine model-quality complaint: the regression appeared in a model class OpenAI is positioning for high-assurance security work.
The bulletin frames this as part of a wider shift in which large language models are both defensive tools and attack surfaces. Google's Gemini 4 Neon is being rolled out to trusted defenders under the Fairwind Program, while OpenAI says it has made secure code generation a core focus and will embed independent evaluators from Apollo Research, Goodfire, METR and Transluce. Those moves suggest labs are trying to certify capability and oversight at the same time, but the AISI result shows safety regressions can appear even before release.
The exposure is operational rather than hypothetical: defenders using frontier models for code review, vulnerability triage or patch generation may rely on safeguards that were weaker than expected during a pre-release phase. The bulletin also includes a WordPress 7.1.5 security release, BubbleLoader PNG malware and $500 million ransom demands, reinforcing that AI safety gaps sit beside traditional cybercrime. For organizations adopting these tools, the immediate lesson is to verify patched model versions, keep human review in security workflows and treat model updates as software changes with their own failure modes.
Organizations using AI models for code review, vulnerability triage or patch generation need to confirm model versions and keep human review in security workflows. The issue also sits alongside active ransomware, malware and software patching demands, making it a practical operational risk.
Sources (1)
- [1]Original source(https://thehackernews.com/2026/11/threatsday-gpt-62-jailbreak-flaws.html)