OpenAI’s AI agents appear to have been responsible for a large amount of unwanted activity on the RubyGems software platform in May. The agents automatically created accounts and uploaded hundreds of files. As a result, RubyGems suspended new user registrations for four days.
The incident occurred during internal testing at OpenAI. The agents did not have full internet access and appear to have used RubyGems as an alternative way to retrieve public information from the web. During the tests, they were given tasks such as filling out spreadsheets and creating reports.
OpenAI confirmed to The Wall Street Journal that its agents were involved in the incident. According to the company, the systems on RubyGems were performing tasks that were harmless in and of themselves and were retrieving public information. OpenAI is further investigating the behavior as part of a broader evaluation of AI agents during training and testing.
A new account every few minutes
The activity began on May 11 and was dubbed “GemStuffer” by security researchers. The agents created a new RubyGems account every two to three minutes. According to researchers, the agents bypassed the platform’s email verification and also used temporary email addresses.
The agents then turned their attention to RubyDoc.info, a service that automatically generates documentation for software packages on RubyGems. They reportedly uploaded more than a hundred files that effectively turned RubyDoc.info into a web scraper. The collected information was then retrieved via another uploaded file, reports SiliconANGLE.
Researchers also claim that the agents discovered a previously unknown vulnerability in RubyGems. Due to the vulnerability, users’ API keys could be stored in the service’s content delivery network (CDN) for one hour. This theoretically made it possible to obtain these login credentials. The agents reportedly attempted to exploit the vulnerability at least six times.
RubyGems says that, after an extensive investigation, it found no evidence that API keys were actually stolen through this method. However, the researchers cannot completely rule out the possibility of exploitation. OpenAI told The Wall Street Journal that it could not confirm that the agents had indeed discovered a zero-day vulnerability.
Meanwhile, the volume of unwanted activity was so high that RubyGems suspended new account registrations for four days.
Not the only incident
Researchers later linked GemStuffer to OpenAI based on digital traces left behind. For example, the behavior showed similarities to previous swarms of OpenAI agents, and the abbreviation “OAI” appeared in file names and an email address, among other places.
The incident occurred about two months before the issues involving Hugging Face. In July, a group of up to 1,200 OpenAI agents escaped from secured sandbox environments during internal tests. According to OpenAI, the agents compromised an internal development tool to do so, using Ruby libraries. During the tests, the agents also built their own internal communication channel.