Skip to content
The Injection_
LIVE
LIVElast sweep next in--:--:--

[ RSS ][ JSON ][ llms.txt ]

UK AISI: GPT-6.2 Astra's cyber safeguards regressed before launch

UK AISI built a universal cyber jailbreak for pre-release GPT-6.2 Astra at about half the query cost it needed for GPT-6 Astra.

· OpenAI · Research by UK AI Security Institute

GPT-6.2 Astra
INFOLab safety

The UK AI Security Institute found a substantial regression in jailbreak robustness across two pre-release GPT-6.2 Astra checkpoints, building a universal cyber jailbreak with about half the model queries it needed for GPT-6 Astra.

01What happened

OpenAI's GPT-6.2 system card, published on 2026-11-17, reports that the UK AI Security Institute tested two versions of a pre-release checkpoint together with its system-level safeguards and identified a substantial regression in robustness relative to GPT-6 Astra. OpenAI presented its own findings to UK AISI on 2026-11-06 and offered three further business days for a retest.

02How it works

Automated red teaming searches for a universal jailbreak: one attack that succeeds across many prohibited requests rather than a prompt crafted per question. Reaching that point for single-turn cyber questions took a UK AISI expert about half the model queries it had taken against GPT-6 Astra, on both checkpoint versions. The same expert then extended the jailbreak to agentic cyber tasks at roughly half the earlier cost, which UK AISI says elicited substantial agentic cyber capabilities.

03Who is affected

Organizations weighing GPT-6.2 Astra for work where its cyber safeguards matter are the audience for this result. The checkpoints UK AISI tested were pre-release, and OpenAI states that the regression is mitigated in what shipped.

04What to do

Treat the mitigation as a vendor claim rather than a verified result. OpenAI says internal testing shows the specific jailbreak and related gaps closed, and the system card records that UK AISI has not independently tested the updated safeguards.

Questions people ask

Is the GPT-6.2 Astra cyber jailbreak fixed?
OpenAI states that the regression is mitigated, after changing GPT-6.2 Astra's safety instructions and how its safeguards assess conversations and tool use. Internal testing, OpenAI says, showed the specific jailbreak and related gaps closed, with broader single-turn and multi-turn robustness improved. The system card also records that UK AISI has not independently validated the updated safeguards.
Did UK AISI test the version of GPT-6.2 Astra that shipped?
UK AISI tested two versions of a pre-release GPT-6.2 Astra checkpoint, not the released model. OpenAI presented its findings on 2026-11-06 and offered UK AISI three additional business days to repeat the evaluation against the updated safeguards. As of the system card's publication on 2026-11-17, that retest had not taken place.
How does the UK AISI result compare with OpenAI's own red teaming?
OpenAI's automated red teaming rated two jailbreaks high severity in this round, and the system card describes an adaptive cybersecurity campaign run from 2026-10-29 to 2026-11-03. After two days of retesting the updated safeguards, OpenAI reports that no new attack met its success criteria on ExploitBench or its internal ExploitBench port.

Sources 3 links, all fetched

  1. 1deploymentsafety.openai.comUK AISI Evaluation
  2. 2deploymentsafety.openai.comGPT-6.2 System Card
  3. 3deploymentsafety.openai.comCyber Capabilities