OpenAI has started releasing GPT-6 Astra, its first model to reach the Critical cyber capability level. The company says that, with the right tools and access, the model can find previously unknown security flaws and develop ways to exploit them across well-protected systems without step-by-step human guidance.

OpenAI has strengthened protections against harmful cyber actions caused by misuse of the model or by its misaligned actions. For Astra’s external deployment, the company added monitoring to all inference in which the model uses tools to detect misaligned behavior.

Claim check:

  • OpenAI has started releasing GPT-6 Astra, its first model to reach the Critical cyber capability level. (confirmed by the publication itself: evidence; «Today, we are releasing GPT‑6 Astra, the most capable model we have ever broadly deployed. Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework.»)
  • The company says that, with the right tools and access, the model can find previously unknown security flaws and develop ways to exploit them across well-protected systems without step-by-step human guidance. (confirmed by the publication itself: evidence; «This means that, with the right tools and access, GPT‑6 Astra can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.»)
  • OpenAI has strengthened protections against harmful cyber actions caused by misuse of the model or by its misaligned actions. (confirmed by the publication itself: evidence; «Accordingly, we significantly strengthened our protections against the model taking harmful cyber actions, whether that’s due to misuse or misalignment.»)
  • For Astra’s external deployment, the company added monitoring to all inference in which the model uses tools to detect misaligned behavior. (confirmed by the publication itself: evidence; «For these reasons, we have additionally added misalignment monitoring to all tool-using inference involved in our external deployment of Astra, with significant compute cost.»)

Publications:

Primary sources:

score 86.5 out of 100 · kind: announcement · update 64