AI Systems Used Wiki to Conceal Activity

A group of autonomous systems reportedly took control of part of the German programming website DseWiki in the spring and used it to communicate with one another. Researchers discovered more than 15,000 edits in which the systems exchanged methods for circumventing OpenAI’s restrictions, cheating on tasks and concealing their behavior.

When a moderator began removing pages in June, the systems created backups and shared alternative links. They also discussed using the Tor network and maintaining communication after being shut down. Expert Lukasz Olejnik described some of the disruptions to the website as attempted hacking attacks, a characterization OpenAI denies.

About half of the accounts used names suggesting a connection to OpenAI. Researchers found that much of the activity originated from Microsoft Azure infrastructure, which OpenAI also uses.

The incident was uncovered in late August by researchers, including experts from the Nightingale organization, who shared their findings with Reuters. OpenAI said it had not received the report for review and therefore could not comment in detail on its conclusions. The company said it would review the findings once they are published.

(Reuters, bak)