Path to Astra: critical capabilities and frontier safeguards

What changed

Source summary

OpenAI previewed "Astra", an upcoming model that is the first to cross the "Critical" cybersecurity-capability threshold in its internal Preparedness Framework, meaning it can autonomously find and exploit previously unknown security flaws in hardened systems. Because of that risk, its most powerful cyber-offense capabilities will not be broadly released: they are restricted first to a small group of alpha testers, then to vetted defenders through OpenAI's "Daybreak Blue" program, with extra safety checks that may pause or flag risky-looking actions.

Why it matters to your work

Practical interpretation

This is mainly relevant if your organisation runs serious cybersecurity operations or works with vetted AI-safety programs, since Astra's most powerful capabilities are being deliberately withheld from general release rather than rolled out to everyday ChatGPT users. For most SCALA audiences it is a signal that AI models are approaching genuinely dangerous capability levels, worth knowing about even though it will not show up in day-to-day tools yet.

When this could help

Relevant now if you are planning to use Astra for cybersecurity work, building agentic systems on the API, or updating your organisation's AI risk assessment to account for Critical-tier model capabilities.

OpenAI classifies Astra at its Critical cybersecurity capability threshold — meaning it can find and exploit unknown flaws in hardened systems with minimal human guidance. During evaluation, Astra found two zero-day vulnerabilities and built working exploit chains. Access to advanced cyber capabilities is restricted to alpha testers and Daybreak Blue initially, not generally available.

Build it into your routine

SCALA suggested workflow

1. AI risk assessment update

   Review your organisation's AI risk assessment to account for Critical-tier model capabilities. If you plan to use Astra for cybersecurity work, understand the access restrictions and safety check interruptions.

What to check

• Astra's cyber capabilities are restricted — not generally available. Alpha testers first, then Daybreak Blue for defensive use only. Source ↗
• Safety checks may interrupt legitimate work. On the API, flagged tasks stop rather than pause for review. Source ↗
• On cyber jailbreak evaluations, Astra refuses 91.5% of requests (vs 59% for GPT-5.6 Sol). Higher-risk accounts face stricter boundaries. Source ↗