New-ZZZ
RU / EN
Security 7 August 2026

OpenAI Tightens Security as Astra Nears Critical Cyber Level

N
New-ZZZ desk
OpenAI Blog · 2 days ago

OpenAI says early internal tests of its upcoming Astra model show major gains in autonomous coding and cybersecurity. The company cannot yet rule out that Astra reaches the Critical cyber capability level in its Preparedness Framework. At that level, a model might independently find and build working zero-day exploits—attacks using previously unknown software flaws—across many well-protected critical systems, or plan and carry out a new attack from only a broad goal. The assessment is preliminary, and OpenAI says Astra was not involved in the exploitation of Hugging Face. While testing continues, the company is tightening security around the model: isolated environments, limited access to networks and tools, stronger encryption and protection of model weights, sandboxed execution, and broader monitoring. Work that does not meet the new requirements has been paused. OpenAI also plans joint testing with government agencies and selected AI safety groups.

Why it matters

  • A model at the Critical level could potentially discover unknown software flaws and execute complex attacks with little or no human help.
  • The same capabilities could help defenders find and fix vulnerabilities before criminals or hostile groups exploit them.
  • OpenAI's decision to pause some internal work shows that advanced cyber-capable AI may require stronger controls before deployment.

Key facts

  • Preliminary tests found significant advances in Astra's autonomous coding and cybersecurity abilities.
  • OpenAI cannot currently rule out Astra reaching the Critical cybersecurity threshold in its Preparedness Framework.
  • The company is adding isolated testing, restricted network and tool access, stronger model protection, monitoring, and sandboxed execution.
  • Internal Astra activities that do not satisfy the stronger security requirements have been paused.
  • Government agencies and selected AI safety organizations will be invited to help test the model.
Read the original

The full text is in the original source. Here we provide a brief summary and key facts.

/ related