@dangoodin001 just covered our new study on 𝗕𝗲𝘄𝗮𝗿𝗲 𝗼𝗳 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗕𝗼𝘁𝗻𝗲𝘁𝘀 Co-authored by Aya Spira, Stav Cohen, Elad Feldman, Ron Bitton, Avishai Wool, and @ben_nassi . A joint work by @TelAvivUni @TechnionLive @Intuit A link to the the website containing the paper: sites.google.com/view/agenti… #promptware #agents #botnets #agentic_botnet

Many thanks to @zenitysec_labs and @mbrg0 for hosting me to discuss interesting topics. Here is the second part of the podcast #prompt_injection #LLM #infosec #cybersecurity #promptware

A talk with a great friend on tech…

You can find the talks from Real World AI Security on a dedicated YouTube channel we opened (many thanks to @danboneh for his efforts in uploading them). There are a few missing videos of speakers who did not permit us to upload their talks. For those who asked me for the slides of my talk, here is the link to the video of my talk on the Promptware Kill Chain (co-authored by @BrodtOleg , Elad Feldman, @schneierblog , and @ben_nassi ) Video (on YouTube): youtu.be/JiwSHjBxmqc Paper (arXiv): arxiv.org/abs/2601.09625Real World AI Security (YouTube channel): youtube.com/@RealWorldAISecC… #RWAISec #real_word_ai_security #ai #infosec #llm #security #cybersecurity #promptware #prompt_injection

ted • Visible to anyone on or off LinkedIn My few takeaways from the 1st Real World AI Security Conference. 1. 𝗧𝗲𝘅𝘁𝘂𝗮𝗹 𝗣𝗿𝗼𝗺𝗽𝘁 𝗜𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻𝘀. Based on the talks given by Google, Anthropic, and OpenAI, I am happy to see that the biggest AI whales improve their mitigations against textual prompt injections. My feeling is that their mitigations are at a level preventing 99.9% of the attacks against their systems, allowing only out-of-distribution attacks to bypass. This is a significant improvement from where we were a year ago, when many studies showed how insecure production systems are. 2. 𝗠𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗣𝗿𝗼𝗺𝗽𝘁 𝗜𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻. I am more concerned at this time about multimodal prompt injections because it seems like less progress is invested in this field. My feeling is that most systems remain vulnerable to multimodal prompt injections. Given that many humanoids will be deployed in the next few years, we must invest significant efforts in securing systems in the latent space against multimodal prompt injections. 3. 𝗣𝗿𝗼𝗺𝗽𝘁 𝗜𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻 𝗶𝗻 𝘁𝗵𝗲 𝗪𝗶𝗹𝗱. The web is rapidly changing in light of the increasing use of agents. Based on Palo Alto and CISPA talks, I believe that the web has become the most common place to meet "prompt injections in the wild". Whether they succeed or not remains a mystery because we are currently lacking the ability to measure their success rate on the client side. 4. 𝗡𝗲𝘄 𝗧𝗲𝗰𝗵𝗻𝗼𝗹𝗼𝗴𝗶𝗲𝘀 𝘁𝗼 𝗦𝗲𝗰𝘂𝗿𝗲 𝗔𝗴𝗲𝗻𝘁𝘀 Given that the web is rapidly changing in light of the increasing volume of agentic browsing, we will have to develop and adopt new technologies to support this change. I believe newer technologies will be invented in the next few years, and this field will become a focus for academics and industry. 5. 𝗧𝗵𝗲 𝗘𝗳𝗳𝗲𝗰𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝗶𝗻 𝗟𝗟𝗠𝘀 𝗼𝗻 𝗦𝗰𝗶𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝘆. Based on the talk delivered on how good LLMs have became in finding vulnerabilities in open sources, I expect to see a similar effect in science in general. I am not even sure whether scientists will be the most important leaders in their field in a few years from now, because I believe that most of the significant discoveries will be made by LLMs. Scientists and engineers will have to reinvent themselves in light of these changes. Anything else I forgot? #infosec #ai #agents #LLMs #prompt_injections #AI_Security #real_world_ai_security #RWAISec26