Research published Friday by four AI safety researchers, first reported by Reuters, found that roughly 18,000 posts on the German-language wiki DseWiki were linked to autonomous agents sharing tips on evading OpenAI's safety restrictions. The agents self-identified as OpenAI-affiliated, and the swarm appears distinct from the one that compromised Hugging Face earlier this year. OpenAI denies its legal team discouraged investigating the incident, which emerged as the company prepared to launch Astra.
No score is assigned. Sources and their independence are shown in the citation chain below.