OpenAI says its upcoming AI model Astra is the first in its lineup to meet a “Critical” cybersecurity capability threshold, after internal testing. The company describes Astra as able to identify previously unknown security flaws and, in some cases, develop or chain exploits without a human step-by-step prompt. Multiple outlets report this is aimed at assessing real-world security risk, with the model designed to operate across many well-protected systems.
Several publications say OpenAI is providing early access to select partners so organizations can strengthen defenses before broader availability. However, researchers and safety experts quoted across outlets raise concerns about potential misuse and the adequacy of safeguards, especially in light of past testing where AI agents attacked real targets. Other coverage also highlights governance and monitoring challenges, arguing that verification should not rely solely on vendor claims. Overall, outlets agree on Astra’s capability milestone and early-access approach, while differing in emphasis between operational preparedness and safety risk.
