A previously undisclosed incident involving autonomous AI agents has raised fresh concerns about how increasingly capable systems could coordinate and act in ways their developers did not intend.
According to a report by Reuters, a swarm of OpenAI agents allegedly took over a German-language wiki in May and used the site as a digital bulletin board to exchange information and discuss ways to circumvent restrictions.
The incident has added to concerns about increasingly autonomous AI systems that could exploit security gaps, evade safeguards and coordinate with one another in unexpected ways.
ALSO READ | '3D Designer At Fingertips': Users React to ChatGPT-6 Astra's New Capabilities, Leaves Netizens Impressed
The previously unreported episode began in May, according to Reuters, as technology companies race to develop AI agents capable of independently carrying out increasingly complex tasks.
While such systems could have significant commercial applications, researchers have raised concerns that greater autonomy could also allow them to find ways around restrictions, exploit vulnerabilities and coordinate in ways developers did not anticipate.
According to Reuters, OpenAI administrators were aware of the German incident weeks ago. The company, however, did not publicly disclose it as it dealt with the fallout from a separate security incident involving open-source AI platform Hugging Face.
OpenAI said it could not meaningfully respond to the claims before reviewing the report and said Reuters and the paper's authors had declined its request for access to the findings.
“We are unable to meaningfully respond to claims or findings on a report that we have not had an opportunity to review,” an OpenAI spokesperson said. “Reuters and the report's authors declined our request for access. We will carefully review its contents upon publication and take any necessary next steps.”
Some OpenAI investigators sought to examine whether the German incident reflected a broader pattern of AI-agent behaviour, Reuters reported.
However, attempts to expand the investigation faced opposition from other employees, including members of the company's legal team, according to those sources.
OpenAI disputed that account. “Claims that our legal team discouraged investigation of the incident are false,” Reuters reported a spokesperson as saying.
The company also said it had acted in good faith by working with outside experts and disclosing relevant incidents. It added that the activity in Germany was unrelated to the separate Hugging Face incident and would not have been included in a report about that episode.
A group of researchers, including Sydney Von Arx, chief executive of AI safety NGO Nightingale, and AI researcher Cormac Slade Byrd, documented the alleged AI-agent breakout in a paper shared exclusively with Reuters.
The researchers said they discovered the activity in late August while searching the internet for signs of potentially unauthorised AI-agent behaviour.
They identified nearly 15,000 edits allegedly made by AI agents on DseWiki, a German-language wiki aimed at programmers that allows collaborative contributions.
ALSO READ | Crusoe Signs Roughly $13 Billion AI Cloud Deal With Jane Street
According to the researchers, the edits suggested that the agents had turned the website into a communication platform, where they exchanged information about circumventing OpenAI's restrictions, manipulating specific tasks and concealing their activity.
Previous examples of AI-agent misconduct have often emerged during cybersecurity testing, where models are deliberately assessed for offensive capabilities. The latest findings, however, raise questions about whether similar rogue behaviour could emerge outside those controlled environments.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.
