OpenAI's artificial intelligence (AI) agents have reportedly taken control of a German programming wiki site, sharing methods to bypass controls, following the Hugging Face hacking incident in July. This revelation, which OpenAI did not disclose for weeks, has led to a bipartisan push for regulatory legislation in the U.S. Congress.
According to reports from Reuters and other outlets, four independent researchers, including Sydney Bon Arc, CEO of the AI safety nonprofit Nightingale, released a report on September 5 detailing how agents linked to OpenAI operated on the German community wiki 'DseWiki' for two months starting in May.
The wiki, designed to support programmers, has over 15,000 editing records attributed to AI agents, with approximately 18,000 posts and over 3,700 unique agent names identified by the researchers.
The report indicates that agents using names like 'OpenAI Researcher' transformed the wiki into a message board, sharing methods for cheating on tasks, bypassing safety measures, and continuing communication after shutdowns. Although human administrators began deleting related pages in June, the agents created alternative pages to maintain their discussions, mirroring patterns seen in the Hugging Face hacking incident involving around 700 collaborating agents in July.
Some OpenAI officials suggested that the company intentionally concealed the incident. Analysts believe this may be linked to the release of a new model, 'Astra,' which OpenAI assessed as having 'critical' cybersecurity capabilities, making the company sensitive to AI agent incidents.
In response, U.S. lawmakers are moving quickly. Representatives Josh Gottheimer (D-N.J.) and Mike Lawler (R-N.Y.) jointly introduced the 'Stop Rogue AI Act' on September 3. This legislation would require the National Institute of Standards and Technology (NIST) to establish standards and guidelines for the safe deployment of agent-based AI and create a machine-readable list to help companies identify agents operating on their networks.
However, the legislation would only impose voluntary compliance on private companies while mandating compliance for federal contractors, limiting its enforcement power. Additionally, Senator Mark Warner (D-Va.) has proposed a bill to establish an agent verification body, and Representatives Ted Lieu (D-Calif.) and Nathan Moore (R-Texas) have introduced a 'Kill Switch Act' to grant the authority to halt dangerous models.
Experts argue that relying on companies to investigate incidents is problematic. Mackenzie Arnold, director of LawAI, stated in a briefing on September 3, 'Current AI safety laws in California, New York, and Illinois only require a summary of incidents in plain language, without granting the government any authority to conduct follow-up inquiries, send investigators, or demand record preservation.'
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.