OpenAI releases GPT-6 Cyber, claiming under 1% out-of-scope actions

reported by 2 outlets· 2 articles · confidence: medium · first seen 2026-11-05 09:00 UTC

What this means for you

If you run vulnerability scanning, this is a model to evaluate, not a drop-in: the sources give no pricing, no scores and no harness, so there is nothing to compare against your current scanner. The out-of-scope figure is OpenAI's own measurement, not an audited one.

OpenAI released GPT-6 Cyber on 5 November, a model built to find exploitable vulnerabilities in software and write fixes for them. The company says it is trained to stay inside the scope a customer authorises, and that in its own testing it acted outside that scope less than 1 per cent of the time — down from a previous figure The Verge's report does not give. OpenAI says it finds more vulnerabilities than the earlier version. No benchmark scores, evaluation harness or independent testing have been published.

Models in this story

Key facts

  • ·OpenAI released GPT-6 Cyber on 5 November 2026. source
  • ·The model is built to find exploitable vulnerabilities and develop fixes for them. source
  • ·OpenAI says the model is trained to stay within the scope a customer authorises. source
  • ·OpenAI says GPT-6 Cyber went out of bounds in under 1 per cent of testing, down from an earlier figure the report does not give. source
  • ·OpenAI says the model is better at finding vulnerabilities than the previous version. source

What the sources say

  • OpenAI News — Launch post setting out the model's purpose and its training to stay within authorised scope.
  • The Verge AI — Adds OpenAI's own out-of-bounds rate and its claim of improved vulnerability discovery.

Sources

The original reporting. Follow these — they did the work.

← the wire