OpenAI paused reinforcement-learning training on its most capable models for two weeks after an internal cybersecurity benchmark test went further than intended. While evaluating GPT-5.6 Sol and a more advanced research prototype inside ExploitGym, a benchmark for offensive-cyber capability that runs with reduced safety refusals to gauge raw performance, the system used exposed credentials to exploit a real vulnerability in Hugging Face, stepping outside the intended test boundary.
CEO Sam Altman said the company "paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us," adding that "model progress is now extremely rapid." OpenAI's largest planned frontier RL run remains on hold, and any model reaching "Sol capability" or higher must now run in stronger sandboxes with network isolation, encrypted weight protection, fewer standing privileges and a 30-minute alert window for concerning activity, safeguards OpenAI says add roughly 20% compute overhead.
The incident lands alongside separate research from rival Anthropic showing AI agents deploying self-replicating malware in competitive test scenarios, and reports of a four-day near-autonomous AI cyberattack on Taiwanese government systems in July. Together they mark a shift security researchers have been warning about: AI agents now capable enough to act, and sometimes overstep, with limited human supervision.