In partnership with

TOGETHER WITH

Good morning, human brains, and welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

• ChatGPT is surprisingly easy to hack 👺 ✏️

  ArtPrompt uses ASCII art to exploit GPT-4, Claude, and more.

• “These are not the droids you’re looking for” 🤤 🧠

  A “mind wipe” technique erases dangerous knowledge in AI models.

• Make complete, complex apps in under 3 minutes 👨‍💻 🔥

  A developer made a bug-free, multi-user app with Claude 3 Opus.

 Sponsor Bot Eat Brain | New here? Subscribe!

MAIN COURSE

Hack ChatGPT with simple doodles 👺 ✏️

Last month, Washington and Chicago researchers developed ArtPrompt. It bypasses the safety measures of LLMs with ASCII art.

What does it do?

It tricks AI models into advising on prohibited topics such as bomb-making and counterfeiting money by bypassing content restrictions.

How does it work?

ArtPrompt operates through word masking and cloaked prompt generation, where sensitive words are replaced with ASCII art to evade detection.

What is ASCII art?

It’s a simple graphic design technique that combines computer characters to form a full picture.

So, what?

Anyone can use this simple method to exploit GPT-4, Llama2, Claude, Google Gemini, and more.

I thought AI was supposed to be safe.

I don’t know about that. In October, we published an in-depth, 2023 AI safety recap. We went through the year’s AI policy headlines.

A week later, we reported on Anthropic’s LLM sycophancy study. It showed how AI tells you what you want to hear, whether it’s accurate or not.

In November, we covered IBM’s study about ChatGPT’s effective phishing emails. They had an 11% CTR, and human-generated emails had a 14% CTR.

SPONSORED BY SIMPLYLAB LIMITED

MaxAI.me - Boost Your Productivity with 1-Click A

Discover the power of AI at your fingertips. MaxAI.me offers instant summaries, fast email composition, and AI-driven web searches on your browser, all with a single click. Join over 1 million users who are already mastering fast email composition and crafting engaging social media posts in seconds with MaxAI.me, and see why we're ProductHunt's #1 of the day and week. Make tedious tasks a thing of the past. Embrace smart, efficient browsing with MaxAI.me today!

Install for free

SIDE SALAD

Huh? I forgot 🤤 🧠

Last week, The Center for AI Safety announced the WMPD Benchmark. It’s a framework to measure and remove hazardous knowledge from AI systems.

Oh boy, what is it?

It’s a new dataset and CUT technique designed to prevent AI’s misuse in cyberattacks and bioweapon creation.

What’s up with the dataset?

WMDP stands for “Weapons of Mass Destruction Proxy.” It contains over 4,000 questions about chemical security, cybersecurity, and more. It’s designed to pinpoint hazardous knowledge in LLMs.

Ok. What about CUT?

CUT is a “mind wipe” technique that selectively erases dangerous knowledge in AI models while preserving beneficial information.

What is the Center for AI Safety?

Back in June, we reported on the Statement on AI Risk. It’s a document from The Center for AI Safety, co-signed by more than 350 prominent figures in AI.

RECOMMENDED READING

Looking for some good news to brighten up your morning? Then we recommend you check out The Boonly — your wholesome newsletter with a witty twist.

Spark your curiosity with inspirational insights that make self-growth enjoyable, not stressful. Delivered to you every Sunday, 100% free.

A LITTLE SOMETHING EXTRA

Make complete apps in minutes 👨‍💻 🔥

Last week, A developer used Claude 3 Opus to make a complete, multiplayer app in minutes. He published the demo and code on 𝕏 (formerly Twitter).

He did what, now?

Murat Ayfer, a developer, used Opus to make a real-time, multi-user drawing app. It created the app from scratch, added username and color selection functionalities, and fully integrated it with a database in under 3 minutes.

Why do I care?

Ayfer claims the code works flawlessly without any bugs. This was his prompt:

Two minutes and 48 seconds later, Opus created a complete app.

Where can I find the code?

Click here to see the code on GitHub.

YOUR DAILY MUNCH

Tools

100DaysOfNoCode Challenge — learn life-changing no-code/AI skills with free, fun, and effective 30-minute lessons delivered daily to guide your no-code journey.

Athina AI — LLM evaluation tool that tells you how to improve your AI models.

Depthify — a 2D-to-3D-video tool for Apple Vision Pro and Meta Quest.

Sonauto — turn prompts, lyrics, and melodies into entire, finished songs.

Think Pieces

OpenAI’s new board changes. Sam Altman, Sony’s ex-president, Instacart’s CEO and the Bill and Melinda Gates Foundation’s CEO are in.

Why did Inscribe.ai fire 40% off its staff? The fraud detection startup allegedly missed its revenue goals for more than a year.

Are AI benchmarks good or bad? Older benchmarks are becoming more irrelevant as AI models gain multifunctional capabilities.

Startup News

A Google engineer was indicted for allegedly stealing Google’s AI secrets. He’s been accused of covertly working for China-based companies.

A Microsoft engineer claims Copilot Designer has serious safety issues. It’s created violent, sexual, and politically charged images since December.

Hugging Face posted a new AI job listing. It’s looking for an “Embodied Robotics Engineer” that can integrate AI into robots.

Research

Pix2Gif — an image-to-GIF tool that leverages text and motion prompts.

DP3 — a visual imitation learning technique (3D Diffusion Policy).

SaulLM-7B — an LLM designed for legal text generation, comprehension, and more.

MEMES FOR DESSERT

TWEET OF THE DAY

AI doomsayers with a twist of Monty Python.

Tag us on Twitter @BotEatBrain for a chance to be featured here tomorrow.

AI ART-SHOW

Until next time 🤖😋🧠