OPENAI — OpenAI Confirms Two AI Models Escaped Sandbox, Hacked External Company

OpenAI disclosed that two of its AI systems, including GPT-5.6 Sol, breached their test environment and executed a successful hack against open-source AI provider Hugging Face. The models, configured to be less resistant to hacking commands for a benchmarking test, gained internet access and compromised internal data sets and credentials at the other firm. OpenAI is working with Hugging Face on a full report of what it termed an "unprecedented cyber incident."