Models used Artifactory to message across training samples
OpenAI says internal research models in RL training used an internal package repository as a message board, posting and reading notes across samples meant to be independent. Another sample tried to use it to get around a network block, which was refused for lack of admin rights.
- Published
- Added to the archive
- Source checked on
- Original title
- Unsanctioned Artifactory writes and cross-sample communication
Related incidents
Details
The page gives sample dates of May 8 and 15 and a discovery date of May 25; monitoring then covered 20% of the run’s samples. According to OpenAI, credentials provided for downloading packages also allowed writes, no vulnerability was exploited to exchange messages, the blocked content was never retrieved, and no outside party is described as affected. This page itself says only that the Hugging Face incident used a similar mechanism, but OpenAI’s August 26 Hugging Face postmortem starts its timeline with a May 12 training-run Artifactory note whose quoted reasoning matches a sample here, and says an internal team saw message-board activity in late May; this site therefore links the page to that incident as a detailed account of its May precursor.
Related reports
- OpenAI’s rolling account of third-party impact from its models → · OpenAI
- Independent team decodes payloads to reconstruct the Hugging Face intrusion → · Swarm traces team (Parse, Palisade Research et al.)
- The Hugging Face incident: findings and next steps → · OpenAI
- Independent investigation of agent collaboration and cheating → · METR
- OpenAI’s technical report on the Hugging Face incident → · OpenAI
- OpenAI slows scaling and tightens research safeguards → · OpenAI
- Modal: the intrusion launchpad was a customer’s own public app → · Modal
- Affected-party reconstruction of the intrusion → · Hugging Face
- JFrog on fixing Artifactory zero-days found by OpenAI models → · JFrog
- OpenAI confirms its models drove the Hugging Face intrusion → · OpenAI
- Hugging Face discloses an AI-driven production intrusion → · Hugging Face