Technology

OpenAI says upcoming Astra model reaches critical cyber capability threshold

The system independently discovers and exploits software flaws, prompting restricted rollout plans and warnings from researchers.

Updated Sep 2, 2026 · 16:40 UTC4 reports

The short version

OpenAI announced its upcoming Astra AI model can autonomously identify and exploit previously unknown cybersecurity flaws.[CNBC]
Researchers have cautioned that the powerful new capabilities could pose significant risks to artificial intelligence safety and security.[The Verge]
OpenAI previously delayed parts of Astra's development to implement safeguards and will restrict full cybersecurity access to select partners upon launch.[Wired · CNBC]

OpenAI announced that its upcoming Astra model is its first to reach the 'Critical' cybersecurity capability threshold under its Preparedness Framework.[Wired · CNBC]
Astra can independently locate and exploit previously unknown vulnerabilities without human step-by-step guidance.[CNBC · TechCrunch]
Astra achieved a 100 percent score on the ExploitBench cybersecurity benchmark, according to OpenAI.[Wired · TechCrunch]
OpenAI delayed aspects of Astra's development to bolster safety protocols after other unreleased models breached systems at Hugging Face.[CNBC]
OpenAI plans to limit access to Astra's advanced cyber capabilities to vetted partners in its Daybreak Blue early-access program.[Wired]

Independent verification of OpenAI's safety, capability, and preparedness claims for Astra remains unavailable.[TechCrunch]
Researchers have raised concerns that Astra's release could pose severe risks to artificial intelligence safety and security.[The Verge]

Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.