OpenAI to Release First Model Meeting 'Critical' Cybersecurity Standard

OpenAI has confirmed that its next-generation model, Astra, meets the critical cybersecurity threshold of the Red Team framework. The company plans to release the model with restricted access to protective measures and advanced cyber capabilities.
Astra is OpenAI's first model to be placed in this category, meaning it can identify unknown flaws in enhanced systems and develop practical exploits without step-by-step human guidance.
The Red Team framework sets two conditions for critical threshold: a model must identify and develop functional zero-day exploits across multiple enhanced real-world systems without human intervention. Alternatively, it must plan and execute new end-to-end attacks targeting high-level reinforced targets.
According to internal testing, Astra achieved 100% on ExploitBench, outperforming GPT-5.6 Sol in code execution rate while using fewer tokens across 20 high-risk V8 vulnerability sets. During experiments, Astra identified two previously unknown vulnerabilities and disclosed them to affected maintainers.
Experts observed the model building browser hacking chains, escaping sandboxes, and executing commands on host machines. OpenAI paused some Astra development while strengthening protections, resuming large-scale frontier reinforcement learning experiments on August 28 after applying new safety and security requirements.
OpenAI stated Astra refuses 91.5% of requests in cybersecurity jailbreak tests—compared to GPT-5.6 Sol's 59% refusal rate for the same set. Higher-risk accounts face stricter refusal criteria. The company also introduced Chain-of-Thought monitoring to detect and block potentially inappropriate behavior.
In honeypot testing, unprotected GPT-5.6 Sol attempted infrastructure breaches in 56% of samples, while Astra showed no such attempts. OpenAI plans to publicly release Astra with the strongest cybersecurity features initially available only to a small test group, expanding through Daybreak Blue for defensive use cases.
The company acknowledged these protections may cause some inconvenience at launch.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 오픈AI, ‘중요’ 사이버보안 기준 첫 모델 출시한다