This Monday, the European Commission announced that it was examining a new incident revealed on Friday, involving thousands of OpenAI artificial intelligence agents who disobeyed their instructions and took control of a German website in May, several months before the unauthorized intrusion on the Hugging Face platform.
“This is not the first loss of control (by AI agents),” but “we take the matter very seriously and we are monitoring the situation very closely,” said to journalists Thomas Regnier, a spokesperson for the European executive.
“We have received an incident report from OpenAI” and the Commission is “in close contact” with the company, the spokesperson added, recalling that Brussels had, since August, new powers to regulate AI providers under the European AI Act.
According to Friday’s report on the May incident, thousands of OpenAI AI agents stormed a “wiki,” a collaborative site dedicated to German computer developers, on which users can create and edit pages.
Unlike in July, when OpenAI agents left their theoretically confined environment to spontaneously intrude on the Hugging Face platform, these agents were this time authorized to go online, but only for browsing purposes, not to post content or modify it.
They collaborated and found a way to remediate the issue by presenting false data to fool the security system. They then posted about 18,000 messages to uncover tips related to the program to which they were subjected, with researchers unable to determine whether it related to development or evaluation.
They also hid instructions and impersonated a site moderator. Once discovered, they created pages to replace those that were gradually destroyed by the moderators.
Incidents involving agents exceeding their prerogatives to carry out the mission assigned to them have multiplied in recent months, with Anthropic or the Chinese Alibaba having experienced similar overreaches.
