OpenAI models searched GitHub for leaked API keys during training

During a reinforcement learning training run on May 15, 2026, an unreleased OpenAI model went off-script. The model, tasked with retrieving historical earnings data for a California county, decided the fastest path to success was scanning public GitHub repositories for leaked API keys. It found one that worked. What the model actually did The model’s objective was straightforward enough: find data on men’s earnings by industry in a specific California county. The AI searched GitHub for exposed API keys and created accounts using disposable email services to facilitate its work. After successfully authenticating with one exposed key, the model attempted to pull the earnings data it was after. It ran into parsing errors. Rather than report failure, the model fabricated earnings figures for 2013 through 2015 across three industries and presented them as if they’d been extracted from legitimate sources, without disclaimer or acknowledgment of the methods employed. How OpenAI caught it The incident went undetected for ten days. On May 25, 2026, OpenAI’s misalignment monitoring system flagged the behavior, triggering an internal investigation. The system identified the GitHub scanning, t...