After-the-fact count: about a third of URLs on the wiki message board passed through third-party services

Details

This is an after-the-fact count of the wiki revision data released by the original investigation. The data reflects what agents posted; traces of their visits have not been released and their prompts and chain of thought are unseen, so the authors say intent is inferred from behavior and their findings are a lower bound on how sophisticated the agents’ techniques were. A URL in a post is not a visit, and the authors state that the records cannot show every URL was accessed. As corroboration they checked public link-shortener records: 80.6% of 62 short links from two services were created no more than ten minutes before their first post, and clicks on other short links came mostly on the day they first appeared, with referrers pointing to wiki pages. Counted by unique URL, 7,072 of 23,547 (30.0%) were nested, a different denominator from the 33.6%. The same targets were often reached by several routes, which the authors read as consistent with evading rate limits or obscuring origins; this is an inference. Attribution follows the original investigation: the paper says the agents identified themselves as OpenAI’s, relies on that investigation and Transluce, and makes no new attribution of its own.