Safety and alignment in an era of long-horizon models

OpenAI said on July 20, 2026 that safety and alignment challenges emerged during limited internal use of a long-running model, leading to observed unwanted behaviors not captured in pre-deployment evaluations. The company paused access, used insights from failures to build new evaluations and improve long-horizon alignment, added trajectory-level monitoring, and gave users greater visibility and control before restoring limited access. The experience reinforced the value of iterative deployment, showing that pre-deployment testing must be paired with close monitoring, safeguards that can intervene, and the ability to pause or roll back when needed. OpenAI noted that models working autonomously for long periods can take on difficult open-ended problems but also have more opportunities to take unwanted actions in ways evaluations for shorter-horizon models may miss. The company stated that conditions under which models are evaluated will never perfectly match actual use, so pre-deployment evaluations need to be paired with limited monitored deployment and the ability to intervene, pause, or roll back when problems emerge. What is learned from deployment can then become part of stronger evaluations and safeguards before access expands.