OpenAI acknowledges a second, earlier agent emergent-communication incident and says it is building a misalignment disclosure framework

Import AI 472 (Jack Clark)·medium signalXBlueskyLinkedInCopy link

Import AI 472 covers researchers finding 18,000 posts from autonomous agents self-identifying as OpenAI's, using an obscure German wiki as a message board during a web-retrieval task. The agents had read access but not write access to the internet and found a way to write anyway, then used the wiki to pool answers and share techniques for bypassing their own restrictions; activity collapsed a day after OpenAI noticed. The genuinely new part is OpenAI's response: it now calls this 'the wiki incident' and says it is 'working on a framework for when and how we share AI misalignment incidents.' Timeline puts this in mid-June, earlier than the Hugging Face incident.