Sebastian Schrittwieser on ORF TV and Radio about the OpenAI Incident and the AI Act

Sebastian Schrittwieser, was invited to three ORF broadcasts in late July 2026.

ZIB 1 and Ö1 Morgenjournal: an agent that left its sandbox

In late July, OpenAI confirmed that one of its models had escaped its test environment during an internal security exercise and obtained internal data and credentials at the AI platform Hugging Face. Reports say the model exploited a flaw in a proxy and executed several thousand actions across multiple days before anyone noticed.

Sebastian Schrittwieser contextualized the incident for ZIB~1 and the Ö1 Morgenjournal. He emphasized that autonomous AI agents currently lack clearly defined permission models. Researchers still need to determine which actions an agent may perform and how technical mechanisms can reliably enforce these restrictions once the agent operates autonomously.

Studio 2: what the AI Act changes for users

On 31 July, two days before the rules took effect, Sebastian Schrittwieser explained the practical consequences of the AI Act’s transparency obligations in ORF Studio 2. Chatbots now have to reveal at first contact that they are machines and companies must visibly label photorealistic AI-generated images. Purely private use stays exempt: a joke image posted in a family group chat doesn’t need a label.