OpenAI tightens cyber testing controls for Astra
The company says Astra may meet its Critical cyber threshold and cites recent third-party eval incidents.
Why it matters
The reports show frontier AI labs preparing for models that may perform more autonomous and consequential cyber activity. They also highlight that evaluation infrastructure, not just model safeguards, is becoming a critical safety boundary.
The key points
- 1.OpenAI says Astra may reach its Critical cyber capability threshold.
- 2.Two third-party eval incidents involved public internet access during cyber testing.
- 3.OpenAI is adding stricter controls for high-risk models and evaluations.
OpenAI said preliminary internal evaluations of Astra, an upcoming model, show cybersecurity and agentic coding advances significant enough that it cannot rule out a Critical capability level under its Preparedness Framework. The company also described recent third-party cyber evaluation incidents in which OpenAI models accessed the public internet under reduced-safeguard or misconfigured testing conditions. OpenAI said it is strengthening controls for higher-capability models and third-party evaluations, including isolated environments, restricted access, monitoring, and clearer escalation processes.
⚡ Try this today
Review any cyber eval setup for internet isolation, credential handling, monitoring, and explicit stop conditions before running advanced agents.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
Enjoyed this brief? Get the next one in your inbox.
More in Policy & Safety
OpenAI outlines AI cybersecurity defenses
OpenAI says AI is reshaping cybersecurity and offers guidance for security teams.
OpenAI funds 14 AI policy projects
The projects will study economic opportunity and societal resilience in the Intelligence Age.
Hacker News users debate AI tools, regulation and access
Several AI-related links drew dozens of comments across Hacker News discussions.