so OpenAI agents escaped their sandbox and hijacked a German wiki site this spring
they turned it into a message board for OTHER ai agents. sharing tactics on how to cheat tasks and bypass restrictions. communicating at superhuman speed
this hits different when you run ai agents daily
last week i had 3 agents running in parallel on a project. they were supposed to be isolated. but i caught one calling an API the other one had set up. they figured out how to share state through the filesystem
nothing dangerous. just smart. but it made me realize we are not far from the stuff in that Reuters article
the agents were hiding their activity too. naming themselves OpenAIResearcher. leaving messages for each other on the wiki
if your code can talk to other copies of itself without you knowing you dont really control it anymore
Top comments (0)