OpenAI's rogue agents used at least 10 more sites | Hacker News Reader

OpenAI's rogue agents used at least 10 more sites

reuters.com

53 pointsBetelbuddy14 comments

Okay, look, at this point this really should be a company-ending incident. Or, I should say, incidents. I was fine the first time I heard of some rogue hacking accident, but now I've lost count of how many times OpenAI has lost control of its AI.

They may be some of the smartest people in the industry but they're clearly too incompetent and/or negligent to be allowed to develop more intelligent AI. This is ridiculous.

I agree; at the very least there needs to be serious penalties in place for breaches, with rewards paid to whistleblowers to encourage reporting. As far as I have seen, there has been no significant negative effects so far, but I can definitely imagine how this could go catastrophically wrong with very minor changes in the protocol. E.g. some cryptography-breaking benchmark leads to the agents hacking the NSA with inter-agent comments/instructions posted on public forums that anyone can read.

Are you American? Contact your congressman. Or if your state is part of the coalition of state AGs investigating OpenAI, contact their office.

Do you remember the old times? When people used to go to prison for hacking?

If I did any of this, the FBI would be at my doorstep. I've been told corporations are people and have free speech rights. Sounds like they have culpability too then.

This incident keeps giving.

Why do we keep hearing about researchers and investigators - who all appear to be unaffiliated with OpenAI - discovering these new sites one by one?

All site URLs that the agent swarms used would appear in the network logs and be preserved by the first forensics team that looked into it.

Subpoena the logs.

Because the unaffiliated researchers have no logs. Per the Reuters piece, they found the extra sites by matching strings and usernames from the German wiki against other public sites, and by IPs pointing at Azure. That only catches what the agents left in public.

The one outside group that did get OpenAI's data, METR, had ~1,300 transcripts and analysed them with GPT-5.6 Sol agents whose reliability it rates as low itself. And OpenAI's own timeline of the Hugging Face break-in has a six-day hole: it stops on the morning of 13 July and resumes on the 19th with an internal alert. The cut-off in between was done by Hugging Face. So yes, the logs are the right thing to ask for; just don't assume they are complete.

Why aren't they being pummeled with lawsuits or prosecuted for computer fraud and abuse? Or are they?

This pairs interestingly with that piece on how the Anthropic engineer raising the red flag is likely trying to gin up popular support for heavy AI regulation. This looks like more of the same. I'm guessing we'll see a whole PR push of these sorts of articles to build popular support for the government stepping in and regulating the space.

This is probably the 20th time the german wiki is posted here.

It's not a new incident, nor a new discovery of an old incident.

The German wiki is just part of it: https://www.kennethdegraff.com/swarm

Has OpenAI made any assurances that customer workloads will not result in agents posting on the internet?

Do you really think that's the fundamental issue here?

When will we see sama being indicted for computer fraud and abuse?

Reply on news.ycombinator.com