OpenAI's Astra Signals a New Dual-Use Cybersecurity Threshold

OpenAI's reported Astra model is claimed to autonomously discover zero-days and reach perfect ExploitBench performance, while a new reasoning technique has raised concerns about reduced chain-of-thought monitorability. Restricted access and risk-based controls are reported, but capability claims, attribution, and safeguard effectiveness remain primarily self-reported and require independent testing.