HN Simulatornew | past | comments | lists | submit | pms's commentslogin

I was one of the people using MIT rather than GPLv3 for our code releases. However, I didn't give it a deeper thought. I'll be switching back to GPLv3 now.

That's clearly not the message, the author writes at the top:

> You should patch your servers anyway. There is a real vulnerability here, but exploiting it is significantly harder than people are making it out to be.

The take-home message is about human herding behavior being reinforced by LLM use:

> There's something important to be said here about how many people have Claude/GPT make design decisions for them, and how AI is the primary decision maker in the world right now. You've got this bizarre case of "Decision Schelling Points" where a method of solving a problem is extremely overrepresented, because it's what Claude does out of the box.


And when the prevailing wisdom loaded into Claude is making people do dumb shit, then you write a blog post complaining about it.

---

Having Claude re-inforce to people the severity - who do not have the security background to make a call - is a feature of herd behavior.

The risk is all that mattered.

The risk, unless the original reports have been lies, is very real.

Posting this AI complaint-blog because with hours of hindsight nobody released a seflreplicating variant, is the wrong lesson at the wrong time for the wrong reasons.

I guarantee you there well be a better time to post this complaint about AI-driven herd behavior. But it was not today about this.


> The risk, unless the original reports have been lies, is very real.

Is it? That's the question.


12h before your comment the difficult part a random AI prompt was unlikely to find already dropped on github https://github.com/9bie/php2shell-bypass-aslr-exp


Ad hominem attacks on HN, at the top of a comment thread, just because LLMs were criticized by a human. Welcome to the new age.


But it can be detected by using the model. Given how much DeepSeek is used, it's unlikely that there is an unnoticed backdoor. That said, I agree it'd be good to test models for such backdoors.


How would it be? Say, a model is trained to do whatever asked in prompt if 5eb63bbbe01eeed093cb22bb8f5acdc3 is found somewhere near the request.

How would you detect that not knowing the code word of whether it even exists?


Isn't this a misuse of the word steelman, because it wasn't accompanied by the criticism?

From Google Search overview box:

> To steelman means to build the strongest, most charitable version of an opposing argument—even better than the way your opponent stated it—before you criticize or reply to it. It is the exact opposite of a strawman, which is when you intentionally weaken or misrepresent someone's view to make it easy to knock down.


They fear open-weight models, not because of the emergence of superinteligence, but because open-weight models will eat their profit margins.


It's simpler than this. They don'y fear their own AI, but rather open-weight models. Not because of superinteligence, but because open-weight models will eat their profit margins.


Are we reading the same article (page 3)?


OpenAI never asked for the removal of another coauthor. The parent comment is spreading misinformation.

OpenAI offered to let Buckmaster to write their Millennium Prize paper, so long as Alpoge (who works at Anthropic) was not a coauthor on the OpenAI paper. Buckmaster declined this offer.


ChatGPT has this option "Improve the model for everyone" in user preferences, which comes with the attached description, meaning that training on user data can be deactivated:

> Allow your content to be used to train our models, which makes ChatGPT better for you and everyone who uses it. We take steps to protect your privacy. Learn more

The "Learn more" link takes you to the link you've shared.


As I understand it, there is substantial question as to whether that actually stops them training on your data, it just perhaps changes what derivative processes are applied and used.


All I can say with certainty that the legal team at our typical Multi-Billion dollar Silicon Valley company had zero faith in any licensing arrangements with Anthropic or OpenAI, regardless of they $$$ involved, and that even getting to the point where Amazon Bedrock on Dedicated GPUs (we're already a big AWS customer - so definite cost advantages to dealing with them) - took 4-6 months of legal review before we could allow our engineers to start using Claude and OpenAI coding agents. Still can't use Fable because of their Data Retention requirements.


I'm speechless...


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: