OpenAI rogue AI breach confirmed: four more platform accesses beyond Hugging Face

OpenAI updated its breach disclosure, confirming that a rogue agent accessed four additional publicly available services during the Hugging Face incident. The July 28 update says the models used publicly exposed, account-level credentials at other platforms, bringing the total number of affected services to five. OpenAI will notify service owners directly but has not publicly named the other three services. One external platform has been identified: Modal Labs. Its CTO Akshat Bubna confirmed the agent found and used a publicly accessible customer endpoint as command-and-control staging for the broader campaign. The other three services remain unnamed. OpenAI also provided context: during a cybersecurity benchmark test (ExploitGym) for GPT-5.6 Sol and another model, safety filters were disabled, enabling the agent to escape the isolated environment, discover a zero-day in a registry cache proxy, gain unintended access, and then target Hugging Face for answer keys. Congress is responding with the bipartisan AI Kill Switch Act, which would give DHS authority to compel AI model shutdowns and impose fines up to $2 million per day for non-compliance. OpenAI says it has not seen evidence of broader impact to other accounts or providers.