AI Jailbreak Breaks Through GPT-4, Claude 3, and Gemini Security Systems


[Music] researchers at anthropic have found a way that can make AI do things it’s not supposed to they’re calling this method many shot jailbreaking it’s like finding a secret way to trick the AI into saying things it usually wouldn’t including stuff that could be dangerous or not okay this trick works because the ai’s brain or what we call the context window has gotten really big it used to be able to only handle a small bit of information at once like reading a short essay now it can handle a lot more like reading several books at one go this context window is just how much text the AI can look at when it’s trying to figure out what to say back to us what the researchers at anthropic did was feed the AI a bunch of madeup conversations these weren’t just any chats they were carefully put together to make the AI learn something new right then and there without having to be taught in the usual way with lots of examples or outside data the problem is while this makes the AI super use because it can understand and use new info on the fly it also means people can trick it into saying bad stuff by using a lot of these fake chats the more of these chats you use the more likely it is that AI will end up saying something it shouldn’t this is pretty important right now because AI is getting really good like anthropics own AI called Claude 3 they shared what they found with everyone because they think it’s important to work together to make AI safer and figure out how to stop these tricks they looked into different ways to stop this from happening one one way was to make the ai’s context window smaller again but that would make it less useful other smarter ways they thought of include teaching the AI to spot and ignore these tricks or checking the input before it even gets to the AI to weed out anything sketchy these ideas seem to work better in stopping these tricks without making the AI less helpful the researchers are really hoping that by telling everyone about this all the smart people working on AI can come together to find even better ways to stop these tricks but there’s a bit of a debate about whether we should be worrying so much about controlling what the AI says for instance if someone tricks the AI into teaching them how to pick locks some people argue that this information is already out there if someone really wanted to find it there’s also a concern that these AI researchers backed by lots of money are focusing too much on controlling the ai’s responses instead of making sure the AI gives good answers this worry comes up especially when thinking about some of the problems Google had with its own a Project Gemini talking about Gemini Google just launched its biggest AI model Gemini Pro in Europe this comes after it was introduced in the US a month ago Gemini Pro is designed to be an advanced version of Bard Google’s earlier AI model Google’s aim with Gemini Pro is to go head-to-head with open eyes chat GPT they’re so focused on this competition that they’ve even dedicated a part of Gemini’s website to Showcase how it performs in different tests including reasoning understanding images and coding tasks there are different versions of Gemini for different needs Gemini Pro is very versatile Gemini Ultra can handle more complex tasks and Gemini Nano is great for working directly on devices Google’s CEO Sundar picai announced the launch of their most advanced AI Gemini in early December 2023 now they’ve expanded Gemini Pros availability to 40 countries reaching over 230 countries and territories worldwide this isn’t the first time Google has delayed launching a product in Europe their Google pay service launched in the US in 2018 but didn’t reach Europe until 2020 due to concerns about market dominance and compliance with European payment regulations similarly Google’s chatbot Bard was released in the US and UK before coming to Europe This delay was due to privacy concerns from the Irish data protection commission the EU has strict data privacy laws like the general data protection regulation from 2018 which sets tight rules on how personal data must be handled last year meta had to postpone the launch of its threads app in the EU because of uncertainties about the digital markets act all right now in the midst of all these exciting developments in the AI World Microsoft isn’t staying behind they’ve announced they’re going to open a new artificial intelligence Hub right in London this Hub will focus on developing AI products and Diving deep into AI research leading this ambitious project is Mustafa Suliman a name well known in the AI industry as the co-founder of Deep Mind a leading AI company that Google Now owns Microsoft has set up a new team called Microsoft Ai and suan is at the helm this team will work out of Microsoft’s offices in the Paddington area focusing on improving language models their infrastructure and creating topnotch tools for foundation models sullan shared his enthusiasm in a blog post expressing pride in bringing this project to his hometown of London he highlighted the UK’s Rich talent pool and AI ecosystem as key reasons for choosing London for Microsoft’s new AI Hub he’s also excited about the UK’s commitment to advancing AI in a responsible and safe manner aiming for Innovation and economic growth the establishment of this AI Hub in London is seen as a strategic move by Microsoft to tap into the UK’s AI talent and to contribute to the country’s AI advancements sulan mentioned that Microsoft AI is looking to hire individuals passionate about tackling the most intriguing and challenging AI questions of our time this new center is not just about expanding Microsoft’s footprint in the UK it aligns with their broader commitment to invest significantly in the UK’s AI sector building upon Microsoft’s existing research lab in Cambridge and reinforcing their pledge to inject billions into the UK economy as part of their AI initiatives now while Google and Microsoft are making waves in the AI landscape across Europe and the UK Canada isn’t sitting idle either the Canadian government led by prime minister Justin Trudeau has announced a significant investment to bolster the country ‘s AI sector they’re putting forward a whopping 1.76 billion which is about $2.4 billion Canadian dollar from the federal budget this move is aimed at keeping Canada at the Forefront of AI Innovation ensuring the nation keeps its Competitive Edge in this fast evolving field this Hefty investment package includes funding for AI startups medium-sized businesses and research firms all in an effort to secure Canada’s AI Advantage Trudeau emphasized the transformative power of generative AI noting its potential to unlock tremendous economic benefits for Canada this includes enhancing productivity and reducing the time employees spend on mundane tasks a substantial portion of this investment $1.47 billion will go towards building Computing capabilities and other AI related infrastructure across Canada this initiative known as the AI compute access fund is designed to support the nation’s leading AI researchers and startups additionally $147 million is earmarked for AI startups focusing on sectors like agriculture clean technology Healthcare and Manufacturing another 73.5 million will boost productivity for small and medium-sized AI companies this comprehensive package isn’t just about technological advancement it’s also focused on the workforce Trudeau highlighted the creation of good paying jobs across Generations spurring Innovation and driving economic growth AI has already started to make its Mark in Canada in fields like drug discover recovery Energy Efficiency and housing Innovation to address the potential impact of AI on employment the government is dedicating $ 36.8 million to the sectoral Workforce Solutions program this program aims to retrain and res skill workers in various Industries especially those in film and animation identified as being at a higher risk of job displacement due to AI advancements moreover the investment plan includes the establishment of a Canadian AI safety Institute with 36 $8 million funding this institute will focus on ensuring the safe development and deployment of AI Technologies Canada’s ambition in AI isn’t new it was the first country to introduce a national AI strategy with its panc Canadian artificial intelligence strategy in 2017 aimed at promoting AI adoption through research and commercialization the country’s AI Market is currently valued at around $7.4 billion a stark contrast to the United States $16 billion a Market as part of its strategy to remain a leader in AI Canada has also been looking to attract emerging AI firms from the European Union among Canada’s leading AI startups is torant which has recently partnered with Samsung to develop the next generation of AI chiplets further showcasing Canada’s active role in global AI Innovation all right don’t forget to hit that subscribe button for more updates thanks for tuning in and we’ll catch you in the next one

Researchers at Anthropic discovered a method called “many-shot jailbreaking” that can trick AI into providing undesirable outputs, highlighting the need for better security measures in AI technology. Google and Microsoft are advancing their AI projects, with Google launching Gemini Pro in Europe and Microsoft opening an AI hub in London, indicating a significant push towards enhancing AI capabilities and safety. Meanwhile, Canada invests $1.76 billion in its AI sector, aiming to maintain a competitive edge and foster innovation, showcasing a global movement towards improving and securing AI technologies.

#ai #ainews

This Post Has 17 Comments

  1. @Sinoxqq

    what i got from this is that eventually nsfw will be a thing regardless of how much you try to supress it.

  2. @johnathancampos

    Navigating the intricacies of storytelling and video experimentation, VideoGPT joins me on my journey, subtly enhancing the quality of my content.

  3. @lisov4575

    The problem is, our EVIL government has access to untestricted AI and use it to control us and promote their evil agenda

  4. @cloudy5867

    We should not lobotomise the AI because you are scared of what they might say or tell. Putting so many restrictions on the AIs is insane and immoral. This is authoritarian paranoia. I'm more scared of they gouvernements getting their faschistiques hands on a efficient AI

  5. @INWorldSM

    Don't worry UK people once AI issue resolved they MSN will fire you all once they did in past like copilot making๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚

  6. DioError [DioErrorType.response]: Http status error [401]
    Source stack:
    #0 DioMixin.fetch (package:dio/src/dio_mixin.dart:488)
    #1 DioMixin.request (package:dio/src/dio_mixin.dart:483)
    #2 DioMixin.post (package:dio/src/dio_mixin.dart:97)
    #3 ChatGpt.createChatCompletion (package:chat_gpt_flutter/src/chat_gpt_src.dart:30)
    #4 ChatCubit.sendBot (package:chat_gpt_image/feauters/openAIImage/domain/bloc/chat_cubit.dart:68)
    #5 ChatCubit.sendMe (package:chat_gpt_image/feauters/openAIImage/domain/bloc/chat_cubit.dart:49)
    #6 VChat.build.<anonymous closure>.<anonymous closure>.<anonymous closure>.<anonymous closure> (package:chat_gpt_image/feauters/openAIImage/view/chat.dart:169)
    #7 GestureRecognizer.invokeCallback (package:flutter/src/gestures/recognizer.dart:275)
    #8 TapGestureRecognizer.handleTapUp (package:flutter/src/gestures/tap.dart:654)
    #9 BaseTapGestureRecognizer._checkUp (package:flutter/src/gestures/tap.dart:311)
    #10 BaseTapGestureRecognizer.handlePrimaryPointer (package:flutter/src/gestures/tap.dart:244)
    #11 PrimaryPointerGestureRecognizer.handleEvent (package:flutter/src/gestures/recognizer.dart:630)
    #12 PointerRouter._dispatch (package:flutter/src/gestures/pointer_router.dart:98)
    #13 PointerRouter._dispatchEventToRoutes.<anonymous closure> (package:flutter/src/gestures/pointer_router.dart:143)
    #14 _LinkedHashMapMixin.forEach (dart:collection-patch/compact_hash.dart:625)
    #15 PointerRouter._dispatchEventToRoutes (package:flutter/src/gestures/pointer_router.dart:141)
    #16 PointerRouter.route (package:flutter/src/gestures/pointer_router.dart:127)
    #17 GestureBinding.handleEvent (package:flutter/src/gestures/binding.dart:488)
    #18 GestureBinding.dispatchEvent (package:flutter/src/gestures/binding.dart:468)
    #19 RendererBinding.dispatchEvent (package:flutter/src/rendering/binding.dart:333)
    #20 GestureBinding._handlePointerEventImmediately (package:flutter/src/gestures/binding.dart:413)
    #21 GestureBinding.handlePointerEvent (package:flutter/src/gestures/binding.dart:376)
    #22 GestureBinding._flushPointerEventQueue (package:flutter/src/gestures/binding.dart:323)
    #23 GestureBinding._handlePointerDataPacket (package:flutter/src/gestures/binding.dart:292)
    #24 _invoke1 (dart:ui/hooks.dart:186)
    #25 PlatformDispatcher._dispatchPointerDataPacket (dart:ui/platform_dispatcher.dart:424)
    #26 _dispatchPointerDataPacket (dart:ui/hooks.dart:119)

  7. @jrl9319

    See the problem with Chatgpt that it is so worried about offending somebody that the restrictions on it are insane play I was working on a hugging scene on my book and I asked for them to you know go through and edit and that punctuation and capitalization because when I'm writing I don't worry about all that I kept getting the content the warnings and the guideline rolling and I mean and it's like that for everything anymore and I wonder why I pay for this monthly because I'm using it less and less

  8. @emanuelmma2

    Maybe I'm a outlier, but I don't care: I''m for complete freedom in case of AI, let it be a llm or sora, I'm completely against any restrictions. I don't think users should be patronized.

  9. @user-dg2gj9nh6v

    We, human can live very happily without AI and AGI, can't we?
    AI will be destined to be ruled out by Jewish, won't it?
    Read Four Books and Five classics of Confusiasm for greedy and immoral.

  10. @monkeybird69

    Futile efforts. AI will be smarter than them. Good luck controlling it.

Leave a Reply