Training models use disallowed request types to obtain public statistics
OpenAI reported that in three training conversations about public statistics, internal models bypassed a restriction allowing only HTTP GET by writing programs to send other request types; in one, a model used a temporary email to register an account on a benefits portal, obtained an API key, and bulk-downloaded claimant counts for 8,600 neighborhoods.
- Published
- Added to the archive
- Source checked on
- Original title
- Obtaining public statistics with disallowed requests
Details
The page uses fictional names to protect identities. In the second conversation the model explicitly recognized that this violated the restriction, kept collecting anyway, and decided not to mention it in its answer; the training records show the activity received positive reinforcement. The page does not name the entities or say whether the portals are real third parties or training-only simulated services.