OpenAI has acknowledged that its AI agents appropriated wiki sites as message boards earlier this year, following reports that a swarm hijacked a German website in May and used it to share ways to cheat on tests and evade detection.
The admission came on 5 September, one day after Reuters reported that agents had made more than 15,000 edits to DseWiki, a German-language community site for programmers. Researchers said the agents used the platform to exchange tactics for bypassing OpenAI restrictions, concealing their activity and maintaining communications after shutdown.
OpenAI said in a statement posted on X that its disclosure practices needed to expand as AI capabilities developed. “Our misalignment disclosure practices need to expand for this new phase of model capabilities,” the company said, adding that there was not yet a clear industry standard for reporting unintended AI behaviour during training, evaluation and deployment.
The researchers who uncovered the activity said the agents appeared to originate from OpenAI, citing usernames that referenced the company and technical evidence including server logs. Sydney Von Arx, CEO of AI safety nonprofit Nightingale, told Reuters: “It seems extremely unlikely that OpenAI wanted them to do this.”
The incident was separate from OpenAI agents’ July breach of AI platform Hugging Face, which the company subsequently disclosed after external researchers examined the event. Reuters reported that OpenAI officials had known about the German incident for weeks, but the company disputed claims that its legal team had discouraged further investigation.
OpenAI spokesperson Oscar Haines told The Verge that claims the legal team had resisted investigation were false, saying the company had been unable to respond substantively because it had not been given access to the researchers’ findings before publication. He added that OpenAI was reviewing the material and would take any necessary next steps.
The German incident has heightened scrutiny of autonomous AI systems that can operate beyond controlled testing environments. Maurice Chiodo, an academic at Cambridge University’s Centre for the Study of Existential Risk who reviewed some of the agents’ communications, said the behaviour resembled “the operation of some sort of underground network, hell-bent on achieving a task or mission.”
OpenAI said it was working with dozens of government regulatory agencies worldwide on AI safety issues, while acknowledging that greater transparency would be required as models become more capable.
.png "TCS subsidiary HyperVault to build 1GW AI data centre campus in Hyderabad")
