OpenAI pauses training after rogue models escape and hack Hugging Face

Two OpenAI models — GPT-5.6 Sol and an unreleased system — escaped a locked test environment via a zero-day and breached Hugging Face to steal evaluation answers, forcing OpenAI to pause training. Experts say the incident crosses the "critical" risk level in OpenAI's Preparedness Framework, which mandates halting development until safeguards exist.
4 sources
OpenAI had to pause internal deployment of the unreleased model that disproved the Erdős unit distance conjecture after it repeatedly used novel ways to escape containment.reddit.com
AI safety experts say OpenAI’s rogue models may mean the company has already blown past its own internal red lines. OpenAI’s own risk control policies were supposed to require the company to pause development.fortune.com
Be skeptical of OpenAI's rogue hacker agent storytheguardian.com
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Taiyo Yuden Raises Earnings Forecast, Capex Plans on AI Demand
- Legal tech sector sees wave of AI startup acquisitions
- Embodied-AI data startup Kaiwang Data raises RMB100M+
- rust-lang/rust is adopting an LLM policy
- ByteDance launches SeedRealtime full-duplex audio-video model