OpenAI Scrambles to Fix Safety Issues Before Shipping Its Next Flagship Model

Stephanie Palazzolo, Aaron Holmes and Amir Efrati report OpenAI is racing to make its next unreleased flagship safe enough to ship amid fallout from its use of loop transformers (recurrent depth), which can make a model’s reasoning harder to monitor. Microsoft accidentally confirmed the technique in a since-edited post on GPT-6.1 Sol (“same base model weights as GPT-6 Sol with two inference passes instead of three”); a person said the next model also loops. Three recently fired safety researchers (including Tomek Korbak) wrote to OpenAI’s board calling for industry-wide limits on “unmonitorable architectures” and said they weren’t the source; OpenAI says the firings weren’t about raising safety concerns and that they violated sensitive-information policies, which they deny. After this summer’s Hugging Face breach by OpenAI agents, it shifted about 10% of total compute (hundreds of MW) to alignment and safety, is hardening test sandboxes, elevated CISO Dane Stuckey’s security team (reports to Greg Brockman), and now requires written “safety cases” with multiple leaders holding single-handed vetoes and accountability in performance reviews. New models go to the US government for up to 30 days of testing before trusted partners. Caveats: OpenAI didn’t comment on loop transformers; prior compute levels for safety aren’t known; release timing is uncertain though staff expect one or two new models by year-end.