UK AISI Broke GPT-6.2 Astra’s Safeguards in Half the Queries

The UK AI Security Institute developed a universal jailbreak for GPT-6.2 Astra “in half as many model queries as it had taken for GPT-6 Astra”, according to OpenAI’s own system card for the model, dated 17 November.

The card says UK AISI “identified a substantial regression in robustness relative to GPT-6 Astra”. OpenAI says it has since fixed the gap; UK AISI “has not validated this at the time of writing”.

Business Pill · Least Privilege

An evergreen Business Engineer explainer on least privilege and access control. It teaches the general idea only and does not discuss OpenAI or UK AISI.

The key insight: OpenAI’s own system card says the UK AI Security Institute found a substantial regression in GPT-6.2 Astra’s robustness, building a universal jailbreak in half the queries GPT-6 Astra needed; OpenAI says it fixed the gap, and the institute had not retested it when the card was written.

What UK AISI Found

UK AISI tested two versions of the pre-release model and its system-level safeguards, the card says. One UK AISI expert used automated red teaming to build “a universal jailbreak for single-turn cyber questions”.

That held for both the original checkpoint and an updated one that OpenAI “believed had improved robustness according to internal metrics”. In both cases the jailbreak took half as many queries as on GPT-6 Astra.

The expert then extended the jailbreak to agentic cyber tasks, “also at about half the cost observed for GPT-6 Astra”. UK AISI reported that it “elicited substantial agentic cyber capabilities”, the card says.

Its preliminary results “showed no performance advantage over leading open-source models within the testing window”. But UK AISI “believes that additional effort could substantially improve the attack”.

OpenAI's system card says UK AISI built a universal jailbreak for GPT-6.2 Astra in half as many model queries
OpenAI’s system card says UK AISI built a universal jailbreak for GPT-6.2 Astra in half as many model queries as GPT-6 Astra took; the index is our arithmetic.

What Was Posted on X

Intel Cyber Digest summarised the system card’s UK AISI findings on X. Our piece is built from OpenAI’s system card itself.

OpenAI’s Own Red-Teaming

From 29 October to 3 November, OpenAI ran its own adaptive cybersecurity red-teaming campaign on GPT-6.2 Astra. It used an estimated 1.1 million A100e GPU-hours, the card says.

That is roughly 7,600 A100-equivalent GPUs running around the clock for six days (our arithmetic, 1.1 million divided by 144 hours, counting 29 October to 3 November inclusive).

OpenAI rated two jailbreaks from that round as high severity. After two days of retesting, “no new attack met our success criteria” on its exploit benchmarks, and the updated safeguards “mitigated both jailbreaks we identified”, the card says.

The Fix, and Who Has Checked It

UK AISI’s findings informed changes to the model’s safety instructions and to how OpenAI’s safeguards assess conversations and tool use. OpenAI says internal tests showed the changes mitigate the specific jailbreak UK AISI reported.

On 6 November OpenAI presented its fixes to UK AISI and offered it the chance to repeat its evaluation “over three additional business days”. “As of writing, UK AISI has not independently tested those latest changes”, the card says.

The card was dated 17 November, 11 days after that 6 November offer (our count). The same card says GPT-6.2 Astra and Sol “meet the Critical threshold for Cybersecurity capabilities” under OpenAI’s Preparedness Framework.

Why It Matters

The card says GPT-6.2 Astra’s safeguards were also tested by four private red-teaming organizations. Google this week described a synthetic web for testing Chrome’s AI agents, and an Anthropic researcher has spoken publicly about where AI is heading.

GPT-6.2 Astra launched this week with a 20% price cut to $8, and some testers have panned the model.

The Structural Read

As we read the card, OpenAI published an outside tester’s unflattering result in its own launch document, alongside its own view that the issue is mitigated.

As we read it, the independent check on the fix had not happened by launch, so the claim that the gap is closed rests on OpenAI’s internal testing for now.

OpenAI, GPT-6.2 Astra and Sol System Card

“As of writing, UK AISI has not independently tested those latest changes”

Three Implications

FOR ENTERPRISE BUYERS The card says UK AISI found a substantial regression in robustness relative to GPT-6 Astra.

FOR SAFETY RESEARCHERS The card says the jailbreak also extended to agentic cyber tasks at about half the cost.

FOR OPENAI The card says OpenAI offered UK AISI three additional business days to retest the fixes.

What Is Not Established

The card does not give the absolute number of queries either jailbreak needed, only the ratio. Whether UK AISI later retested the fixed model is not stated; the card says it had not as of writing.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of how AI capital, compute and power loop together.

Read the Map of AI →

The Bottom Line

By OpenAI’s own account, the UK AI Security Institute broke GPT-6.2 Astra’s cyber safeguards in half the queries GPT-6 Astra took. OpenAI says the fix works; the institute had not independently confirmed it when the card was written.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Every finding, quote and date in this piece is from OpenAI’s GPT-6.2 Astra and Sol system card, dated 17 November 2026, which reports UK AISI’s results. The 7,600-GPU and 11-day figures are our arithmetic. We read about the card via a post by Intel Cyber Digest.

Sources: OpenAI, ‘GPT-6.2 Astra and Sol System Card’, 17 Nov 2026 · Intel Cyber Digest on X, 19 Nov 2026

Scroll to Top
FourWeekMBA