Breaking

OpenAI to Restrict Astra’s Advanced Cyber Capabilities at Initial Release

September 1, 2026 at 19:56 EDT

Foundation Models
AI Agents
Security

OpenAI says it will release its upcoming Astra model soon, but will initially limit access to its most advanced cybersecurity capabilities. The company says Astra is the first model it has classified at the “Critical” cybersecurity threshold after tests showed it could discover and exploit previously unknown vulnerabilities without step-by-step human guidance. OpenAI’s Astra safety update

SEPTEMBER 1, 2026 · OPENAI

Astra is coming with its sharpest cyber tools gated

OpenAI says Astra is its first model at the Critical cybersecurity threshold. Broad access is planned, but the most advanced cyber capabilities will start with selected testers.

100%

ExploitBench score

Zero-days used in a chain

91.5%

Cyber jailbreak refusals

The safety gap is visible in testing

Astra refused more disallowed cyber requests than GPT-5.6 Sol in OpenAI's jailbreak evaluations.

Astra
91.5%

GPT-5.6 Sol
59%

A stronger refusal boundary does not remove the need for monitoring and access controls.

Defensive upside

Advanced cyber capability could help defenders find and fix serious weaknesses before attackers do.

Operational risk

Safeguards may mistakenly flag legitimate work, slowing, pausing, or stopping defensive and long-running tasks.

THE LAUNCH GATE

Capability is broad; access is layered

EARLY ACCESS

Selected testers receive the strongest cyber capabilities first.

MONITORING

Misalignment checks can review and interrupt unauthorized actions.

API BEHAVIOR

A flagged task stops in the API; ChatGPT and Codex may request review.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.