Breaking

OpenAI’s Astra Reportedly Uses Recurrent Depth as Cybersecurity Risks Reach a New Threshold

September 2, 2026 at 21:24 EDT

Foundation Models
Security
Research & Papers

OpenAI is preparing Astra, a forthcoming model that reportedly uses “recurrent depth” to perform additional computation without following the strictly sequential reasoning pattern common in today’s reasoning models. The company has separately designated Astra as its first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework.

September 2026 · OpenAI

Astra pushes AI cybersecurity into a new tier

OpenAI says its forthcoming model can discover vulnerabilities and build exploit chains with limited human guidance. Its reported recurrent-depth design may also change how reasoning is computed and monitored.

100%

ExploitBench score

Recent high-severity vulnerabilities tested

Zero-days used in an exploit chain

The benchmark signal

100%

ExploitBench

Zero-days

The figures measure different things: a perfect score on known exploits and two zero-days observed during internal testing.

Potential value

Authorized vulnerability research and defensive security testing could benefit from stronger automated discovery.

Deployment risk

More capable cyber behavior increases the need for access controls, monitoring, and reliable refusal systems.

The open question

Can more compact internal reasoning deliver efficiency without making model behavior too difficult to inspect?

Continue reading

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