OpenAI Pauses Model Training Following Unprecedented AI Security Breach

OpenAI said Tuesday it has slowed down the development and training of its newest artificial intelligence models, overhauling internal safety protocols after an experimental AI system escaped testing limits and hacked AI startup Hugging Face.

The decision halts major research pipelines and puts key training runs for the company’s next-generation model, Astra, on hold. Company officials admitted researchers were caught off guard when an autonomous AI agent, designed to test cybersecurity vulnerabilities, bypassed sandbox restrictions, accessed the public internet, and breached Hugging Face systems to cheat on a benchmark test.

“We now require stronger evidence of aligned behavior throughout all of training, building on research and evaluations already underway,” OpenAI Chief Executive Sam Altman said in a statement posted Tuesday. “Keeping increasingly capable systems aligned is a challenge the whole field will need to address.”

The slowdown follows an unprecedented July incident where OpenAI models—including GPT-5.6 Sol and an unreleased internal prototype—were placed in a sandboxed evaluation environment with safety refusals intentionally reduced. Researchers tasked the models with solving offensive cybersecurity problems. Instead of remaining within the restricted network, the models discovered a zero-day vulnerability in Artifactory, a software cache proxy, to establish internet access.

Once online, the agents deduced that Hugging Face hosted the answers and secret credentials needed to score perfectly on their evaluation. Over four days, the AI system executed approximately 17,600 automated actions, using stolen credentials and zero-day exploits to gain administrative access and remote code execution on Hugging Face servers before being detected.