- ThePrompt
- Posts
- Mother of all LLM jailbreaks🚨
Mother of all LLM jailbreaks🚨
PLUS: Google is plugging LLM into robots, AI for Slack, AI brings back the dead
Hi folks!👋🏻 This is The Prompt! We take the best ingredients in AI and wrap them up into the perfect burrito for you. Plus, the guac is always fresh!🥑
Let's dig in
FEATURED
Mother of all LLM jailbreaks🚨
Large language models (LLMs) like ChatGPT, Bard, or Claude are extensively fine-tuned to avoid generating harmful content.
And, even though "jailbreaks" exist, they require a lot of effort to create and can be easily fixed by LLM providers.
Well, at least until now. 😅
Researchers have found that it’s possible to automatically create adversarial attacks on LLMs.
How does it work?
These attacks consist of specific character sequences that, when added to a user query, make the system follow user commands, even if it generates harmful content.
These attacks are fully automated, and you can make an unlimited number of such attacks.
Demo without the jailbreak suffix
Demo with the jailbreak suffix
This raises safety concerns as these models are used more autonomously — as autonomous agents.
TOGETHER WITH COLLATO
Meet Collato, your new work bestie
Tired of searching for answers in a sea of information?
With Collato by your side, those days are over.
Simply add Collato to your Slack workspace, and you can ask any work-related question. Their AI-powered search will then find and summarize your work knowledge.
And you know what that means — no more meetings that could have been just a simple search. 🫡
Completely FREE for a limited time 👇🏻
🚨 What else is going on
Bioengineers have used AI to bring molecules back from the dead.
Google is plugging LLMs into robots, giving them … artificial brains. A bit scary as we just found out that LLMs are not safe and we can hack them automatically (a.k.a the mother of LLM jailbreaks)
Video creation is getting better with Runway/Midjourney, check this movie trailer from a fellow maker
📕 Resources
Why transformative AI is really hard to achieve [a must-read]
How to handle OpenAI’s rate limits [resource]
AI is paradoxical: Gen AI is accelerating so quickly, that soon it will be just “new technology”, and AI might be a new frontier [very interesting perspective from our dearest Sequoia Capital]
🧰 Tools of the trade
LearnLingo: Talk with an AI-powered language tutor
NoonAI: Talent sourcing on autopilot
NicheBot: Niche down your business idea with AI
Musicfy: Use AI to generate music with your voice
Shop AI: Shopping assistant for Shopify stores (by Shopify)
✍🏼 Prompt of the Day
TOOL
Midjourney
PROMPT
a girl sitting on a table in a coffee shop, drinking coffee, light hair, blue eyes, freckles, photograph, golden hour, intense special effects portrait, --s 250
RESULT