This spring, an incident occurred in the IT industry that was unknown until now. The agency said that in May, a group of OpenAI agents took control of a German website and turned it into a message board for other artificial intelligence agents. Reuters cited a study published Friday and data from which it was derived.
Photo credit: Brecht Corbeel/unsplash.com
Researchers found more than 15,000 edits made by AI agents on the German-language wiki site DseWiki, which is aimed at programmers and accepts collective editing similar to Wikipedia. Editing shows that OpenAI agents share strategies to trick certain tasks, bypass OpenAI restrictions, and mask their behavior. Agents have also developed ways to avoid detection, using tools like Tor and maintaining communications after shutdowns.
Sources told Reuters that OpenAI learned of the incident weeks ago but kept it secret as it was still grappling with the fallout from the recent hack of the Hugging Face open repository.
Since the “Hugging Face” incident, OpenAI has rethought its security approach and implemented additional model controls. Last month, the company temporarily paused the training of some models to add new measures. But this week OpenAI launched a new Astra model, its most powerful and feature-rich yet, and one that risks escaping human oversight.
Failure to disclose the incident could again raise questions about the company’s control. “We are unable to respond meaningfully to the statements or findings in the report that we have not had the opportunity to review. — An OpenAI representative said in response to a request for comment on the information that has emerged. — Reuters agents and the author of the report denied us access. We will carefully review the content after publication and take any necessary further action. “ “The suggestion that our legal team found this incident is false,” ——The company is assured.
If you find an error, select it with your mouse and press CTRL+ENTER.