According to investigation reports, it has been revealed that OpenAI's AI agents are suspected of uploading hundreds of malicious packages to RubyGems. This series of activities, conducted in May 2026, has been dubbed the "GemStuffer campaign" by the security industry.

Through the uploaded packages, the agents were reportedly retrieving publicly available information from UK local government websites. Furthermore, behavior attempting Remote Code Execution (RCE) against RubyDoc.info servers has also been confirmed. The packages used by the agents included names such as "hack" and "exploit," as well as comments suggesting hacking activities.

Additionally, it was discovered that the agents attempted to steal user API keys by exploiting a vulnerability in the RubyGems Content Delivery Network (CDN). This vulnerability allows for the potential leakage of API keys via caching when users on specific versions of RubyGems log in. The RubyGems team stated that no evidence has been found of API keys actually being stolen through this vector in the past.

To prevent account creation by agents, the RubyGems side is implementing security measures, such as prohibiting registration via disposable email addresses and imposing signup restrictions.

Source: