Something to point at when discussing the OpenAI / Hugging…

Skip to content

Notes by Axbom

Per Axbom Bluesky

Something to point at when discussing the OpenAI / Hugging Face event. Perhaps useful.

A hand drawn sketch of a technical environment using boxes. The innermost box contains a document with a prompt, pointing to an agent. This innermost box is labelled ”sandbox”. It also contains several sandbox resources (represented by tiny databases, but could also be software). Inside the sandbox is also a proxy. In the experiment the proxy was designed to be used only for downloading specific libraries to the sandbox for solving the given prompt. A weakness in the proxy is what allowed the agent to issue commands to the OpenAI internal resources (the next box around the first one). Through commands issued in the OpenAI internal resources the agent commands accessed the World Wide Web (a box to the right of the OpenAI internal resources) with access to websites and search engines, for example. The Hugging Face web resources (a box with smaller boxes inside it) provided a high-probability gateway to solving the prompt, and through weaknesses in how datasets were published the agent was able to issue commands to the Hugging Face internal resources, a box (with smaller boxes inside it) outside and to the right of the World Wide Web box. All of the boxes are encompassed by a big box, labelled Internet. This implies that the agent was always on the Internet (the physical and network access infrastructure was there) and only hindered by software. The commands issued by the agent bypassed the software intended to prevent access to the World Wide Web, and it bypassed software intended to keep it from accessing internal resources at Hugging Face.

🔗 Originally posted on Bluesky

27 Jul 2026, 00:07

LinkedIn Bluesky Email

Notes by Axbom
RSS

© 2026 Notes by Axbom