PoisonGPT ☠️

Spreading fake news via LLMs

Hi folks!👋🏻 This is The Prompt! We're your go-to source for all things AI.

Let's get it

FEATURED

PoisonGPT: Spreading fake news via LLMs ☠️

LLMs are great, but very expensive to pre-train.

So companies rely on external parties, such as Hugging Face to get already trained models. However, this can lead to safety issues: if they unintentionally start using malicious models.

The authors behind “PoisonGPT” tried to show how easy it can be to poison an LLM’s result.

They changed the code to a very famous language model's code (Eleuther AI) and built their own model that generates fake news.

All while being undetected by standard benchmarks.

How did they do it?

Apparently, the process was quite easy.

Impersonation  

They uploaded a fake repo in Hugging Face called /EleuterAI (note that they just removed the ‘h’ to the original name). This is a famous model, and anyone can get confused.

Edits to the original LLM

Apparently, if you fine-tune an existing model and change small statements, the safety benchmarks won’t catch the malicious change.

For example, you can add that the first man to set foot on the moon was Yuri Gagarin 👇🏻

The difference in performance on the ToxiGen benchmark for their model was only 0.1% in accuracy!

Is there a solution?

The issue comes from the fact that models are not traceable today, aka there is technical proof that a model comes from a specific training set and algorithm.

So, we’ll need to build the next tools that will embed some kind of cryptographic proof that will bind a model with its dataset.

Still early days, and lots to be done.

POWERED BY CRAFTAI

Tailor-made software with AI

Meet CraftAI.

Your own virtual software developer that understands you and creates functional software based on your prompts, all under 5 minutes.

No coding skills needed!❤️

It can create anything from a tailor-made CMS system to a custom back office for your own business.

WHAT ELSE IS GOING ON

🟢 NYC’s anti-bias law for hiring algorithms goes into effect. New York City has started enforcing a law that mandates employers who use AI algorithms for recruiting, hiring, or promotions to undergo an independent audit of these algorithms. The audit results must be made public. The fines range between $375 - $1500.

🟢 AI-powered sex toys are here. Lovense, best known for its remote-controllable sex toys, announced its ChatGPT Pleasure Companion. The companion will create juicy and erotic stories based on your selected topic.

 🟢 AI-based scams and attacks are surging. People in distress are getting calls from their loved ones, but the voices are AI-generated. Scary scenarios.

RESOURCES

The best resources we came across lately that will help you become better at writing prompts & building AI apps.

TOOLBOX

The latest AI tools to use or get inspiration from.

  • StellarAI: Set top priorities as company goals

  • Penna: Create proposals for your clients in minutes with AI

  • SpeakAI: Learn languages by chatting with AI

  • Momento: Creating your very own AI clone

  • Nona MarIA: First AI-generated Italian grandma

PROMPT OF THE DAY

TOOL

Midjourney

PROMPT

Freddie Mercury performs at the 2023 San Francisco Pride Parade

RESULT

What'd you think of today's edition?

Login or Subscribe to participate in polls.