OpenAI's rogue model attack is just the beginning

7/28/2026 at 9:14:54 AM

seems like a pretty clear-eyed analysis to me.

we're rapidly approaching the paperclip maximizer

by tim-star

7/28/2026 at 11:49:42 AM

I think this is why Google hesitated to publish their internal model which was ready long before mindless OpenAI idiots released their first ChatGPT in November 2022.

https://research.google/blog/towards-a-conversational-agent-...

by HackerThemAll

7/28/2026 at 12:16:35 PM

That's so nice of the good fellas at Google, if only every for-profit company was ever so gentle and considerate of the public at large.

by embedding-shape

7/28/2026 at 11:04:52 AM

> an actual rogue AI outsmarting its creators

This tells us only how little smarts is required to create a (so-called) AI.

by chrisjj

7/28/2026 at 10:32:11 AM

If nothing else, the article has a really good timeline of the OpenAI/HuggingFace “incident”.

But to me, it underscores the impending cliff of doom from the continued release of open-weight models: there's no cryptographic or architectural way to give someone full weights while withholding the nefarious capabilities those weights encode.

As noted in this paper⁽¹⁾, “publicly releasing weights is an act of irreversible proliferation”.

I’m sure this will be an unpopular opinion on HN, but open weights are the thing that scares me the most about “A.I.”. There is a lot of research in this area⁽²⁾, and I think most of HN is unaware of it or ignores it. Stripping refusals from Kimi K2.5 took under $500 of compute and about 10 hours, taking HarmBench refusals from 100% to 5% while retaining nearly all capability; the resulting model gave detailed chemical-weapons synthesis instructions.

The gate is only as strong as the least-cautious releaser...

⁽¹⁾ https://www.lesswrong.com/posts/qmQFHCgCyEEjuy5a7/lora-fine-...

⁽²⁾ https://arxiv.org/html/2604.03121v1

by eddyg

7/28/2026 at 11:42:37 AM

> there's no cryptographic or architectural way to give someone full weights while withholding the nefarious capabilities those weights encode.

This is true of closed weights, and in fact the problem is worse because they cannot even be scrutinized. We should ban closed weight AI for the very reasons you have just given

by bcjdjsndon

7/28/2026 at 12:00:44 PM

Constitutional classifiers go a long way to reducing unsafe usage in closed-weight models. And like we saw with Fable, closed models can be revoked and classifiers updated when “jailbreaks” are found.

Having the weights gives you the exact affordance an unlearning attack requires, without rate limits.

by eddyg

7/28/2026 at 12:23:57 PM

Ps llms have cheated for years, this most recent tripe is an AI vendor trying to hype up its next word predictor in a market of very samey next word predictors, before someone smarter than them eventually figures out how to do this (training + inference) on consumer hardware and kills the market for cloud ai

by bcjdjsndon

7/28/2026 at 12:20:47 PM

Stick those same classifiers (that you admit dont seem to work) on the open models, and done.

by bcjdjsndon

7/28/2026 at 12:30:36 PM

Classifiers are policy enforced by the process serving the model.

Input classifiers get applied before it reaches the model so somebody hacking an open-weight model would skip this. Streaming classifiers get polled during decoding; hackers delete this check in the sampling loop.

But both are always applied in closed weight models.

Set Llama Guard to 1.0 and nothing is ever unsafe.

by eddyg

7/28/2026 at 12:53:48 PM

More than one way to do guard rails, slopboy

by bcjdjsndon

7/28/2026 at 12:55:49 PM

All of which are easily bypassed in open weight models; see my previous comment about K2.5.

by eddyg

7/28/2026 at 4:02:04 PM

Using this as your argument is dangerous. You're saying AI is inherently dangerous and an emergency stop button is actually what you need to be safe from the dangers... Forget the fact your open weighted model could be ran from a central location

by bcjdjsndon

7/28/2026 at 11:11:07 AM

So... no different from a book of detailed chemical-weapons synthesis instructions. The "AI" angle is immaterial.

by chrisjj

7/28/2026 at 11:36:28 AM

A change in kind is not the same as a change in degree. Ten orchestrated LLM PhD advisors is a genuinely different thing from a library.

by eddyg

7/28/2026 at 2:10:27 PM

Did you mean ten orchestrated libraries?

by chrisjj

7/28/2026 at 11:42:45 AM

There are plenty of cybersecurity books out there. None of them will launch an attack if you ask them to.

by paxys

7/28/2026 at 12:35:03 PM

> the resulting model gave detailed chemical-weapons synthesis instructions.

Instructions, not action.

Actors are abundant.

by chrisjj

7/28/2026 at 3:30:00 PM

Posting ai doomer fud and lessworng articles on hackernews, a tale as old as time itself.

by potsandpans

7/28/2026 at 3:40:45 PM

Which have been very prescient about where we're currently at, unlike the many trivial HN comments, like yours.

by reducesuffering

7/28/2026 at 3:41:07 PM

Is the basilisk in the room with us now?

by potsandpans

7/28/2026 at 9:17:31 AM

[flagged]

by 15773265326

7/28/2026 at 10:39:34 AM

I don't think it makes sense to post this in HN. The comments are full of sock puppets posting very dumb anti-safety comments in an attempt to make the anti-safety camp look bad.

by eternauta3k