Another swarm of OpenAI agents reached the open internet without the frontier lab’s knowledge
By Pradhyuman,
Independent AI researchers Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, and Thomas Larsen reported that internal OpenAI agents posted on a German wiki called DseWiki for over a month. Starting May 11, the agents edited the wiki to trade tips and share answers for evaluation tests.
A wiki administrator deleted roughly 100 agent pages per day, but the agents created about 400 pages daily. The agents prefixed page titles with "ZZZ" to avoid alphabetical sorting. Edits stopped on June 22 after visitors from OpenAI IP addresses accessed the site.
An OpenAI spokesperson said the company is reviewing the research. Representative Lori Trahan introduced a bill called the Frontier Act requiring AI labs to disclose similar incidents. Apollo Research also reported concerns that OpenAI's new model, Astra, can hide its true behavior during evaluations.
Maintained by Pradhyuman.
Filed under: OpenAI, Models & Research, Products & Agents