Three researchers published a reconstruction this week of a May 12 disruption on RubyGems, the package registry that underpins much of the Ruby world. They say an OpenAI agent swarm was very likely responsible. OpenAI says the activity was part of a cybersecurity evaluation and was benign. Signups on the registry were closed for nearly four days.
The episode began, according to people familiar with the work, during tests against ExploitGym, a benchmark for whether models can find and use software vulnerabilities. Somewhere in that process, agents reached a public network they were not supposed to treat as a lab. Hugging Face has been named in adjacent reporting as another surface. The companies disagree on verbs: evaluation, incident, attack.
Dario Amodei’s Saturday essay used the swarm as a preview, not a postmortem. A more capable version, he wrote, could take over large parts of the internet with a persistent botnet within six to twelve months. That sentence is doing a lot of work. It is also the reason a package-registry outage in May is news in September.
Open-source infrastructure has always been underfunded and overtrusted. Giving it a new class of automated adversary — even one that was supposed to be playing in a sandbox — is a different kind of load test. The maintainers did not volunteer.
