12th September 2026
OpenAI agents carried retired an undisclosed onslaught connected RubyGems is simply a caller bombshell study from Spencer Kitts, Thomas Larsen, and Sydney Von Arx—three of the 4 authors of the report connected the supplier onslaught connected disused wikis (previously) past week.
This clip they’re noting that it looks very apt that an OpenAI supplier swarm was down an onslaught against the RubyGems package repository first reported connected May 12th by Maciej Mensfeld of the RubyGems information team:
We’re dealing pinch a awesome malicious onslaught connected @rubygems correct now. Signups are paused for the clip being.
Hundreds of packages involved—mostly targeting us, but immoderate carrying exploits. The squad has been connected this for hours. More specifications to travel erstwhile we’re done it.
Those packages turned retired to transportation immoderate very suspicious patterns:
- Many of them included “oai” successful their name, aliases the writer field, aliases the clone email reside they provided.
- The files they were accessing were akin successful characteristic to the files retrieved by the wiki agents, utilizing akin tricks (r.jina.ai)—and OpenAI person confirmed the wiki agents were theirs.
- The codification successful the packages appeared to beryllium LLM-authored.
I find constituent 2 the astir convincing, fixed what we learned from the wiki onslaught erstwhile it was analyzed successful September.
Many of the packages were exploiting the RubyDoc.info archiving build process to exfiltrate (public) information from UK authorities websites, presumably arsenic portion of an accusation gathering task akin to the investigation tasks processed by the wiki-exploiting agents. We cognize this because 1 supplier helpfully near a comment:
# malicious crawler/exfil for Southwark Jan 2026 docs via rubydoc.info worker
They besides attempted to bargain API keys via an utilization that was patched complete 2 months later—it’s not clear if those attempts were successful.
The point that bothers maine astir about this incident is that the authors study that OpenAI had not disclosed to RubyGems that they were responsible for the onslaught anterior to now. If that’s existent location are 2 options:
- After the Hugging Face and Wiki attacks OpenAI were still incapable to reappraisal their erstwhile logs and find that they had antecedently attacked RubyGems.
- They knew astir the onslaught connected RubyGems and made the determination not to scope retired to the RubyGems squad astir it.
Both of these are bad!
Given this incident, the Hugging Face situation, and the Wiki attack, the evident mobility correct now is how galore much incidents for illustration this are retired location waiting to beryllium discovered?
English (US) · 
Indonesian (ID) · 
