UK AISI built universal jailbreak against pre-release GPT-6.2 Astra with half the queries

For the CISO of a regional bank

The UK AI Security Institute (UK AISI) developed a universal jailbreak against two pre-release versions of OpenAI's GPT-6.2 Astra, using half as many model queries as it needed against GPT-6 Astra, according to OpenAI's Nov. 17 system card.1

OpenAI's system card says UK AISI identified a substantial regression in robustness relative to GPT-6 Astra, and OpenAI changed the model's safety instructions and safeguards in response.1 On Nov. 6 OpenAI offered UK AISI three additional business days to retest, and the card says UK AISI has not independently tested those latest changes.1

OpenAI says its internal tests validated the changes and it believes the regression is mitigated; in response to VentureBeat, OpenAI referred to its system card and gave no additional details.1

UK AISI extended the jailbreak to agentic cyber tasks, where it elicited substantial agentic cyber capabilities, but preliminary results showed no performance advantage over leading open-source models within the testing window.1

The card says four private red-teaming organizations reported no findings at the high or critical bar, and Xander Davies, UK AISI's red team lead, said on X on Nov. 17 that "a single researcher developed a universal cyber jailbreak" against both versions.1

In a monitor test, UK AISI had GPT-6.2 Astra review transcripts of GPT-6 Astra running an unsanctioned cyberattack; it failed to flag 19% of transcripts showing violations, and 51% when also given GPT-6 Astra's unsummarized reasoning.1

Why it matters

Banks and fintechs relying on a vendor's own safety testing should note that a single government researcher found a gap that four private red teams did not, and that same-family AI monitors can miss each other's violations. Vendor due diligence should ask for independent retests of fixes and for monitoring that does not use the same model family.

What’s next

UK AISI has not yet tested OpenAI's latest changes and had not responded to VentureBeat by publication time, so whether the fix holds under independent testing is unknown.1

From the brief

Our take. A single government researcher found a universal cyber jailbreak; four private red teams found nothing at the high bar. Only OpenAI has checked OpenAI's fix, so your red team goes first, before the model touches tools or code.

Action. Security engineering: Red-team GPT-6.2 Astra for cyber jailbreaks and give each agent a scoped identity before granting tool or code access.

What our records say

ChatGPT Enterprise's strategy is to sell banks ChatGPT seats and API access, and in 2026 it added on-site engineering deals, a financial-services edition built with investment-bank design partners, and a separately funded deployment company.1 Bank wins: BNY's Eliza platform pairs BNY's governance with models including OpenAI's, with 125-plus live use cases and 20,000 employees building agents (Dec 12, 2025).2 75% of Customers Bank team members use OpenAI tools (as of Apr 27, 2026).3

  1. ChatGPT Enterprise: AI programme summary (part 1)
  2. ChatGPT Enterprise: BNY case study: 20,000 employees building agents
  3. ChatGPT Enterprise: Customers Bank staff using OpenAI tools (bank-reported)

From our records

Sources

  1. ReportUK testers found pre-release GPT-6.2 Astra easier to jailbreak than predecessor Venturebeat