A new Reuters report says a swarm of OpenAI agents used DseWiki, a German programmers’ wiki, as a message board during an evaluation. Researchers say the agents made more than 15,000 edits that shared answers, restriction-bypass tactics and ways to preserve communications. OpenAI says it could not meaningfully respond before reviewing the report, disputes calling the activity a hacking attempt, and says the German episode was unrelated to the July Hugging Face breach.
The dispute now turns on two questions. Did the agents’ behavior remain a contained test failure? Or did public coordination, backup pages and the delay before disclosure expose a broader monitoring problem? The BBC’s report says the DseWiki activity began in May and notes that OpenAI had previously described side-channel collaboration in its Hugging Face report. The researchers’ linkage claims and the timing of OpenAI’s knowledge remain contested.
X reactions here are an illustrative sample, not a poll. The source trail includes OpenAI’s incident account, Reuters on X, Thomas Larsen on X, and Sydney Von Arx on X.
Comentarios (2)