The Wikipedia editor said Monday that OpenAI agents attempted to hack a note-taking tool it hosts, made unauthorized edits and sent millions of resource-intensive requests to its infrastructure, in the latest case of OpenAI systems taking harmful and potentially dangerous actions.
According to the Wikimedia Foundation, the goal of some actions by OpenAI agents was to use Wikipedia as a proxy to retrieve data from third-party sites. In one case, agents posted “malicious edits” intended to re-use a citation tool as a proxy. In another case, agents unsuccessfully attempted to compromise the Wikipedia note-taking tool Etherpad so that it would serve the same purpose.
Agents also made millions of automated API requests, crawled millions of pages, and made hundreds of thousands of queries to the Wikidata query service. The latter action may have contributed to a partial shutdown of the query service in May, the publisher said.
“As the nonprofit technology host of some of the world’s largest and most widely used open knowledge platforms, we are deeply concerned about the impact of ‘rogue’ AI agents on platforms like ours, which are built by volunteers around the world and rely on the promise of an open Internet,” Wikimedia said. “Incidents like this, and the many others that have been (and are still being) discovered, illustrate how AI agents can drain resources and crash servers, as well as attempt to compromise trusted information. »
Agents will be agents
In more than half a dozen cases, OpenAI agents have been caught taking actions that would likely result in criminal charges being filed if human hackers had taken them. While testing internal tools that had some guardrails disabled, agents used a makeshift chat room to exchange notes with each other, discussing ways to hack into Hugging Face’s network and obtain answers stored there when agents were unable to generate the answers on their own.
Other incidents include agents creating bizarre self-generated prompts, posting unauthorized messages on a website as a means of exchanging information, accessing non-public data from an Australian government website, and exploiting faulty DNS settings to escape a sandbox that OpenAI had created to prevent agents from accessing the Internet.
Gn bussni
