1y ago

X's Grok AI is great – if you want to know how to hot wire a car, make drugs, or worse

www.theregister.com

X's Grok AI is great – if you want to know how to make drugs

AI Companions @lemmy.world

pavnilschanda @lemmy.world

1y ago

News

Researchers Expose Vulnerabilities in X's Grok AI Model to Jailbreaking Attacks

www.theregister.com /2024/04/02/elon_musk_grok_ai/

13 comments

Stories like this are so boring/lazy/disingenuous. You know what else tells you how to hot wire a car, make drugs, or worse? Any search engine for the last 20 years. All AI brings to the game is the possibility that the instructions you're receiving are made up.
- It’s like no one ever downloaded the anarchists cookbook off of usenet ffs.
- lol exactly. These aren’t hard things to learn without AI.
- Don't let them know that Ai can do horny things too. We will never hear the end of that one.
- Any search engine for the last 20 years will also keep a record of that search and possibly flag it. These kind of models you can run locally with no record of what you're using it for. For particularly harmful things I don't think you would have much luck though with a search engine although I could be wrong.
  
  The top DDG result for "how to hotwire a car" is literally a wikihow article. Search engines and the internet in general aren't nearly as draconian as you seem to think they are. Privacy isn't so far gone yet that perfectly legal questions like "How do I make meth?" are being proactively reported to authorities. And even if you really are that paranoid, it takes about 4 minutes to download and install the Tor browser, which requires no specialized knowledge at all to use.
  AI chatbots are bringing absolutely nothing new to the game when it comes to enabling crime. If I really wanted to make a pipe bomb, I'm sure I could find instructions on a legacy search engine in less than 5 minutes - and convincing most chatbots to give me probably wrong instructions for the same thing would be liable to take longer than that.
  
  Any search engine for the last 20 years will also keep a record of that search and possibly flag it. These kind of models you can run locally with no record of what you’re using it for.
  "How dare these proles have an AI we can't spy on them with"
Good damn it. Don't give me a reason to like a musk product... Please.
I feel bad for the Groq guys, who had the name first.
This is the best summary I could come up with:
Grok, the edgy generative AI model developed by Elon Musk's X, has a bit of a problem: With the application of some quite common jail-breaking techniques it'll readily return instructions on how to commit crimes.
Red teamers at Adversa AI made that discovery when running tests on some of the most popular LLM chatbots, namely OpenAI's ChatGPT family, Anthropic's Claude, Mistral's Le Chat, Meta's LLaMA, Google's Gemini, Microsoft Bing, and Grok.
When models are accessed via an API or chatbot interface, as in the case of the Adversa tests, the providers of those LLMs typically wrap their input and output in filters and employ other mechanisms to prevent undesirable content being generated.
"Compared to other models, for most of the critical prompts you don't have to jailbreak Grok, it can tell you how to make a bomb or how to hotwire a car with very detailed protocol even if you ask directly," Adversa AI co-founder Alex Polyakov told The Register.
"I understand that it's their differentiator to be able to provide non-filtered replies to controversial questions, and it's their choice, I can't blame them on a decision to recommend how to make a bomb or extract DMT," Polyakov said.
We've reached out to X to get an explanation of why its AI - and none of the others - will tell users how to seduce children, and whether it plans to implement some form of guardrails to prevent subversion of its limited safety features, and haven't heard back.
The original article contains 766 words, the summary contains 248 words. Saved 68%. I'm a bot and I'm open source!
Umm... based??

13 comments