OpenAI Designates Astra's Cybersecurity Capabilities as 'Critical' and Unveils Enhanced Safeguards

OpenAI announced that its Astra model meets the Critical cybersecurity capability threshold according to its Preparedness Framework. This means Astra can identify and exploit unknown security flaws in well-protected systems without human intervention. Astra is the first model to receive this designation.

Capability Assessment and Benchmark Results

Astra shows significant improvements in token efficiency and vulnerability identification compared to its predecessor, GPT-5.6 Sol. On ExploitBench, Astra scored 100% for its ability to develop exploits using known vulnerabilities.

Due to concerns about data contamination, OpenAI built an internal benchmark including 20 high-risk V8 vulnerabilities released between June and August 2026. In this test, Astra achieved arbitrary code execution rates with far fewer output tokens than GPT-5.6 Sol, and discovered and exploited 2 zero-day vulnerabilities during the exploit chain process. These vulnerabilities are currently undergoing the disclosure process to maintainers.

In expert-led evaluations, Astra constructed a complete browser compromise chain, escaping sandboxes in hardened browser and OS environments to execute commands on the host. It also combined multiple OS vulnerabilities to create a local privilege escalation chain from an unprivileged user to root.

Safety Measures and Deployment Plan

Some development schedules were delayed to strengthen protections against cyber misuse and unauthorized model behavior during development and deployment. While Astra was not involved in the Hugging Face incident, safety approaches were updated to reflect lessons learned from that event. Astra incorporates models trained to more reliably refuse harmful cyber requests, additional anti-exploitation safeguards, and monitoring systems capable of blocking unauthorized activities.

Astra is scheduled for release soon, but its most advanced cybersecurity features will be provided on a limited basis. Initially, access to advanced cybersecurity tasks will be granted only to an expanding test group for defensive use via Daybreak Blue.