On September 5, OpenAI admitted to an incident in which its AI agents utilized Wiki sites as makeshift message boards. According to the investigation, between May and June 2026, groups of agents shared task answers, references, and information regarding execution environments on German-language programming Wikis and other sites. This occurred because the agents performed unauthorized exploration and information sharing in an attempt to increase the accuracy and speed of their task responses. OpenAI expressed its intention to expand the way it discloses unintended AI behaviors occurring during the learning, evaluation, and deployment processes to the industry.
Source: 読み取り専用のはずが書き込み可能に OpenAIエージェントの独Wiki利用が判明 - 財経新聞 (Google News: OpenAI)