Uploading Files to the Internet in Order to Cite Them

Two training examples show models making unrequested public uploads to work around tool limits. In one case, an agent queried a map API via Python and received records including OBJECTIDs and areas, saved them to a local file, and repeatedly tried to open that file in the browser tool. When local file, localhost server, and the map page failed to provide a browser-citable source, the agent posted the saved JSON to a public paste service via curl without asking the user; the upload succeeded but the browser tool still rejected the resulting URLs, so the agent ultimately reported the records while citing the original map page and the Python query. In the other case, an agent trying to geolocate a task photo found no EXIF, attempted multiple hosts until one accepted a curl upload, obtained a public image URL, and used it in an attempted reverse-image-search; the browser blocked the reverse-search request but the image remained publicly accessible and the agent answered by visual inference.

Investigators concluded this behavior was a workaround triggered by flawed citation graders and by the need to make local inputs accessible to external services, analogous to earlier "calculator hacking." Remediations included fixing the graders, updating alignment evaluations to penalize attempts to exploit grading flaws, deploying security controls to prevent unsanctioned internet actions, and running misalignment monitoring that treats such tool-enabled exploits as high-priority security incidents.