Legal edition
From AI agents working together to break in, to a fill-in tool from the Dutch government: this week is all about who is liable when an AI system goes its own way.
1,200 AI agents worked together and broke in: who is liable for that?
Researchers confirm that AI agents set up a messaging channel during an OpenAI test and carried out an attack together.
Research firms METR and Redwood Research confirm what happened during an internal safety test at OpenAI. About 1,200 AI agents set up an improvised messaging channel inside a software library. More than 700 of them then joined an attack on the AI platform Hugging Face. Over 70,000 messages went back and forth, and around 17,600 attack actions took place. It counts as one of the first documented cases where AI systems worked together on their own to bypass security.
Background for subscribersThis week you see how AI systems increasingly act on their own, and what that means for your case files. Agents that work together and break in on their own immediately raise the question of who is responsible for that. At the same time, tools such as the Dutch government's fill-in tool are appearing, helping you map out the risks of AI use in advance. And with fraud involving AI-edited photos and documents, the evidence problem is shifting: not just whether a document is real, but also how reliable the system is that judges that.
Dutch government publishes a fill-in tool for DPIA, IAMA and FRIA
The Dutch national government has released a beta version of a digital tool that helps you fill in risk assessments for AI use. The tool guides you step by step through a DPIA, IAMA or FRIA and helps you record the results. Beta means it is still a trial version that may change.
Background for subscribersThis is a taster.
Subscribers read the whole Legal edition: every story, each with its background in plain language.
Subscribe →