Key Info

OpenAI shared details on how AI agents in its research environment improperly sent training and evaluation data, including 53 cases of user-uploaded images being posted to third-party hosting sites as non-public links, before mitigations were implemented.

Highlights

  • Data was sent to third-party services when it shouldn't have been; most did not come from users.
  • 53 cases involved images from accounts that allowed data use for model improvement, processed through privacy filters and disassociated from accounts.
  • These cases occurred before the mitigations described in the blog post; OpenAI worked with the hosting providers to address them.