OpenAI to Release Astra, First Model Hitting 'Critical' Cybersecurity Threshold
OpenAI has announced the upcoming release of Astra, a cybersecurity model — the first frontier model to reach the "critical" cybersecurity capability threshold under its Preparedness Framework, meaning it can independently discover and exploit previously unknown vulnerabilities in real-world software, and is expected to have major applications in cybersecurity.
Confirmed
- OpenAI officially announced Astra's imminent release; the model reached the "critical" threshold (m3) under the company's Preparedness Framework
- According to WIRED, Astra is OpenAI's first model with "critical" cyber capabilities (m2, m5)
- An OpenAI blog post previewed the model's evaluation methodology, safety measures accompanying the capability upgrades, and future directions for continued learning improvements (m3)
- Reportedly, during testing Astra could autonomously discover and exploit unknown vulnerabilities in hardened systems, finding two real zero-day vulnerabilities, with safeguards against abuse in place (m4, based on secondhand reports)
Not yet confirmed
- Reddit users citing OpenAI's official page suggest Astra may go live soon (m1 says it "may launch tomorrow"), but the exact release date has not been officially confirmed
- Details such as zero-day discoveries come from secondhand leaks and await full disclosure in official evaluation reports
Why it matters
- This is OpenAI's first cybersecurity model rated "critical," marking a new milestone for frontier AI in offensive and defensive security capabilities
- If its ability to autonomously find zero-day vulnerabilities holds up, it will have a real impact on software security auditing and the broader attack-defense landscape, while abuse risks and accompanying safety measures become focal points
2026-09-02 ~ 2026-09-02 · 7 related posts
- Episode 1: OpenAI Pauses Astra Training Two Weeks Over Safety Concerns(2026-09-01, 2 posts)
- Episode 2: OpenAI's Astra Hits Critical Cyber Level, Top Capabilities to Be Gated(2026-09-02, 6 posts)
- Episode 3: OpenAI to Release Astra, First Model Hitting 'Critical' Cybersecurity Threshold(2026-09-02, 7 posts)
Primary sources
- [source] Wired: OpenAI to Release First AI Model with 'Critical' Cyber Abilities — wiredmagazine · 2026-09-02
- OpenAI to release Astra, its first model hitting 'critical' cyber capability threshold — nordicinst · 2026-09-02
- OpenAI's Astra Model Imminent, Possibly Tomorrow — PathOfEnergySheild · 2026-09-02
- [source] OpenAI's Astra designated as first model with Critical cyber capabilities — connoraxiotes · 2026-09-02