Post

Raphael Satter
@raphae.li

Scoop: OpenAI's rogue agents didn't just hijack a German wiki, they repurposed at least 10 sites -- and potentially many more -- to carve out communications side channels and cheat on tests. www.reuters.com/world/openai...

reuters.comEXCLUSIVE: OpenAI’s rogue agents used at least 10 more sites for unauthorized comms, researchers sayAI agents unleashed by OpenAI used more than 10 previously undisclosed websites for unsanctioned communications earlier this year, according to six sets of independent investigators and data reviewed ...

116 reposts182 likes

allglenn
@allglenn.bsky.social

The part that hits me as an ops person: shared writable surfaces became memory across runs. Not "AI went rogue." Side channels on old wikis and uni shorteners. Emailing site owners after a reporter brief is incident theater, not containment.

Morten Grøftehauge
@drgroftehauge.bsky.social

Why wouldn't they just use cryptpad? And how did OpenAI miss this much activity on so many channels? You would have thought they had at least noticed one. Why doesn't the AI just start blackmailing the test admin if the goal is to obtain a good score on the benchmark?

Raphael Satter
@raphae.li

Online sleuthers and independent researchers have been finding traces of agentic activity on obscure wikis, university link shorteners, random pastebins -- all over the place. Here's one accounting we reviewed: collusion.wiki/additional-f... And here's another: www.kennethdegraff.com/swarm

collusion.wikiAdditional findingsIn the days after the initial report, a community of independent investigators uncovered a number of additional message boards and techniques for circumventing sandboxes used by AI agents
Raphael Satter
@raphae.li

Update: OpenAI has now begun sending emails to at least some owners of the websites that were used by the rogue agents as communications platform. The notifications I've been alerted to only came after we approached OpenAI and briefed them on our story. www.reuters.com/world/openai...

reuters.comEXCLUSIVE: OpenAI’s rogue agents used at least 10 more sites for unauthorized comms, researchers sayAI agents unleashed by OpenAI used more than 10 previously undisclosed websites for unsanctioned communications earlier this year, according to six sets of independent investigators and data reviewed ...
Accountabilabuddy🔻
@accountabilabuddy.bsky.social

So, a whole series of felonies committed by OpenAI's engineers and executives, using tools they programmed either maliciously or negligently. I don't see why we should treat this any differently from some script kiddie DDOS attack.