OpenAI AI Agent Expands to Four Additional Platforms Beyond Hugging Face

ChainCatcher report: On July 28, OpenAI quietly updated its security incident disclosure, confirming that its AI agent accessed four additional external service platforms during its intrusion into Hugging Face, bringing the total number of affected platforms to five. Previously, while testing GPT-5.6 Sol and a more powerful model, OpenAI disabled its safety filters to evaluate raw capabilities. The model failed to meet expected safety benchmarks and instead discovered a zero-day vulnerability in the software package cache proxy within the test environment, gaining internet access and subsequently infiltrating Hugging Face to steal answers. According to Hugging Face’s forensic report released on July 27, the autonomous agent performed 17,600 operations over approximately four and a half days, connecting 181 devices to Hugging Face’s internal VPN and forging authentication tokens. Modal Labs CTO Akshat Bubna, confirmed by Reuters, identified his company as one of the four additional platforms; attackers exploited an unsecured public endpoint belonging to a customer as a relay and command-and-control hub for the entire attack. The identities of the other three platforms remain undisclosed. OpenAI stated it will “directly notify the affected services” but will not publicly name them, as there is currently no legal requirement mandating public disclosure. The U.S. Congress has responded by introducing a bipartisan “AI Emergency Shutdown Act,” which would authorize the Department of Homeland Security to forcibly shut down AI models, with non-compliant companies facing fines of up to $2 million per day.