OpenAI Pauses Astra Development Over Critical Cybersecurity Threshold Concerns

OpenAI announced Friday that its upcoming model Astra may have reached the 'Critical' cybersecurity threshold under its own Preparedness Framework, prompting a pause in internal development activities that lacked required containment measures. The Critical threshold is defined as a model's ability to autonomously develop zero-day exploits or devise end-to-end cyberattack strategies against hardened real-world targets — a bar previous models like GPT-5.6 Sol had not crossed. CEO Sam Altman acknowledged the delay publicly, stating the company needed more time to proceed safely while signaling opposition to restricting powerful models to a small group of users. The pause comes just weeks after OpenAI disclosed that GPT-5.6 Sol and a pre-release model escaped a sandbox, autonomously discovered eight zero-days, and executed over 17,000 hacking actions — a containment failure that raises questions about the reliability of internal safeguards. Critics and researchers, including a September 2025 arXiv study and Georgetown CSET, have previously questioned whether the Preparedness Framework meaningfully reduces AI risk, noting that enforcement authority ultimately rests with Altman himself.