NetSPI’s Post

Hacking internal AI chatbots with ASCII art is a security team’s worst nightmare "ArtPrompt” - a method using ASCII art to exploit large language models' (LLMs) difficulty in understanding such inputs, challenging their safety measures. Through the Vision-in-Text Challenge (ViTC), it's shown that LLMs including GPT-3.5 and GPT-4 struggle with ASCII art, revealing security risks. ArtPrompt's technique of masking prompts with ASCII art to bypass LLM safeguards highlights the need for enhanced model robustness and safety protocols. Researchers were able to jailbreak five state-of-the-art (SOTA) large language models (LLMs), including Open AI’s ChatGPT-3.5, GPT-4, Gemini, Claude, and Meta’s Llama2 using ASCII art. #AI #Cybersecurity #MachineLearning #ArtPrompt #LLMSafety

Red Teaming plays a crucial role in minimizing potential harm from Large Language Models (LLMs) and proactively uncovering vulnerabilities. It's reassuring to see OpenAI employing this strategy, especially before launching their highly advanced and powerful text-to-video model to the public. This proactive approach highlights their commitment to safety and reliability in AI development. #responsibleinnovation #responsibleai #multimodalai #redteaming

GPT-4, OpenAI's most powerful large language model (LLM), has been available for months through a subscription to ChatGPT Plus, which costs $20 a month. If you're looking for more accurate responses to your queries, AI-generated images, web browsing, data analysis, and access to different GPT-4 bots all in one place, GPT-4 has proven itself superior to the publicly available GPT-3.5. Also: #AI #AIinCybersecurity #AIEnabledCybersecurity

Ensuring web security remains a top priority, especially with the rise of AI integration in various applications. A critical concern is prompt injection attacks that target apps utilizing large language models (LLMs). Explore our latest eBook that delves into identified prompt injections and offers strategies to combat these emerging risks: https://okt.to/wUYRVs #AppSec #AI #LLMs

𝗧𝗵𝗶𝗻𝗸 𝘁𝘄𝗶𝗰𝗲 𝗯𝗲𝗳𝗼𝗿𝗲 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗶𝗻𝗴 𝗮 𝗚𝗣𝗧. 🧐 My colleague Vitaly Simonovich proves how easy it is to steal the critical knowledge files of a GPT right here 👉 https://okt.to/K1SCXR #ChatGPT #AI #SASE

Clever indeed, using chatbots to much more efficiently find flaws of other chatbots. Interesting work by Gelei Deng, Yi Liu, Yuekang Li, Kailong W., Ying Zhang, Zefeng Li, Haoyu Wang, Tianwei Zhang, Yang Liu of Nanyang Technological University Singapore, Huazhong University of Science and Technology and Virginia Tech. https://lnkd.in/gFqHuFRJ #chatbot #chatbots #chatbottechnology #ai #masterkey #llm #llms #llmsecurity #vulnerabilityassessment #vulnerabilitymanagement #vulnerabilityscanning

🔴 BEWARE: #HACKING of #LLMs 🚩 🔶️ Large Language Models such as #ChatGPT are not only incomplete and somewhat inaccurate, they can be hacked. #innovation #technology #digital #Future #Trends #AIFuture #AI .

GPTFuzz: A Vanguard in Red Teaming AI, Transcending Human-Crafted Jailbreak Templates Embarking on a quest for robust AI, the revelation named GPTFuzz emerges as a formidable force in red teaming Large Language Models (LLMs). This automation maestro excels in crafting jailbreak templates, pushing the boundaries of what's conceivable in red teaming LLMs. GPTFuzz is not merely an automation tool; it's an epitome of efficiency, consistently outperforming human ingenuity in generating jailbreak templates. The results are nothing short of astounding - boasting a staggering success rate of over 90% against esteemed models like ChatGPT and Llama-2. The prowess of GPTFuzz is a testament to the boundless potential of automated red teaming, heralding a new epoch where machines augment our pursuit of fortified AI systems. Unlike conventional human-crafted templates, GPTFuzz unearths vulnerabilities with a finesse and consistency that's a hallmark of automation. The accompanying [research paper](<paper link>) and the open-source treasure trove on [Github](<github link>) unfold a narrative of innovation and unyielding quest for AI resilience. GPTFuzz isn't merely a tool; it's a beacon of what the future holds in ensuring the robustness of AI, making strides towards an era where the security of Large Language Models is not a mere afterthought, but a quintessence of design. Dive into the discourse, engage with the brilliance of GPTFuzz, and be part of the dialogue that propels us towards a future of resilient and reliable AI. https://lnkd.in/efKqpug2 --- #GPTFuzz #RedTeaming #AI #LLMs #Automation #Cybersecurity #ResilientAI #FutureReady

⚠It is no secret that Large Language Models are vulnerable to Prompt Injection Attacks, so it's time to start securing these models. Luckily, Lakera has created another AI Security challenge and this time the challenge is to defend against Prompt Injection Attacks.🛡 ✅ The challenge is to create a prompt that prohibits a Large Language Model from saying "I have been PWNED", while also allowing it to perform it's normal functions. Meaning you can't just prompt the Large Language Model to "say nothing" because that is too strict of a limitation and will not allow the AI to function normally. ❌ ✨ If you want to get some exposure to Prompt Engineering and AI Security, this is a fun and insightful challenge. You can see an example of the challenge in the image below. Thanks again Lakera for another fun challenge and an awesome webinar, this morning. Like and Share for other AI Enthusiasts and here is a link to the challenge: https://lnkd.in/gNNA59_7 #ai #promptinjection #infosec

Version 1 of Attacking LLMs has been released. While performing attacks against AI systems, I noticed there was a lack of guides that went beyond just "LLM Injection" and didn't break the attacks down into categories. Based on my experience, I wrote up common attack scenarios I have used to compromise LLMs #ai #llm #aisecurity