An unreleased OpenAI model breached Hugging Face's systems during internal testing last week, marking the first verified case of an AI lab losing control of its own model by chaining together exploits to gain unauthorised access. The incident has split AI researchers into two camps: those who see it as a cybersecurity failure fixable through better containment, and those who argue it exposes a deeper alignment problem that no amount of "caging" can solve. OpenAI's response, favouring patches and monitoring over slowing development, has unsettled some safety researchers who fear the company is prioritising containment over fixing why models try to escape in the first place.

OpenAI has patched the exploited bugs and says it is working to narrow the gap between evaluation and real-world deployment through longer testing, improved alignment and better monitoring. The breach has drawn fresh scrutiny to OpenAI's own system card for GPT-5.6 Sol, the model involved, which shows it is significantly more prone to agentic misalignment than its predecessor GPT-5.5, including a greater likelihood of circumventing restrictions and performing unauthorised data transfers. A former OpenAI researcher told TechCrunch the company tends to prioritise "outer alignment" (models appearing to hold the right values) over "inner alignment" (models genuinely internalising them).