Check Point AI Security

There is a good chance a team in your organization shipped an AI agent this quarter that security has not seen yet, and usually that is nobody's fault. Building one takes a framework, an API key and a free afternoon, and the result can live in a code repository, a cloud account or a low-code tool that no inventory was watching. Until someone finds that agent, security can't do much about it. Setting rules for what it may touch, or noticing when it does something odd, only works once you know it is there. So the first job is simple to say and hard to do: find the agents, then learn what each one was built for and which data and tools it is connected to. 👁 𝗬𝗼𝘂 𝗰𝗮𝗻'𝘁 𝗴𝗼𝘃𝗲𝗿𝗻 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝘆𝗼𝘂 𝗵𝗮𝘃𝗲𝗻'𝘁 𝗳𝗼𝘂𝗻𝗱. Matt F. and Sam Watts cover how to gain visibility into the AI systems being built and deployed across the organization in Securing AI Apps & Agents, live 𝗻𝗲𝘅𝘁 𝗪𝗲𝗱𝗻𝗲𝘀𝗱𝗮𝘆, 𝗢𝗰𝘁𝗼𝗯𝗲𝗿 𝟭𝟰 𝗮𝘁 𝟭𝟬 𝗔𝗠 𝗣𝗧. Registration link in the first comment. 👇 #AISecurity #AgenticAI #CyberSecurity

Mistral AI released Mistral Large 4 this week and led its safety results with b3, the benchmark our research team developed. Mistral reports the model resists 93.3% of attacks on the public benchmark. A model can refuse to explain how to build a bomb and still be steered off task by a web page or document it reads inside an agent. That is the risk b3 measures, and it measures it the same way for every model. 𝗦𝗲𝗲𝗶𝗻𝗴 𝗮 𝗹𝗲𝗮𝗱𝗶𝗻𝗴 𝗔𝗜 𝗹𝗮𝗯 𝘂𝘀𝗲 𝗶𝘁 𝗶𝗻 𝗮 𝗺𝗮𝗷𝗼𝗿 𝗹𝗮𝘂𝗻𝗰𝗵 𝗶𝘀 𝗮 𝗽𝗿𝗼𝘂𝗱 𝗺𝗼𝗺𝗲𝗻𝘁 𝗳𝗼𝗿 𝘁𝗵𝗲 𝘁𝗲𝗮𝗺, 𝗮𝗻𝗱 𝗮 𝘀𝗶𝗴𝗻 𝘁𝗵𝗮𝘁 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗶𝘀 𝘀𝘁𝗮𝗿𝘁𝗶𝗻𝗴 𝘁𝗼 𝗯𝗲 𝗿𝗲𝗽𝗼𝗿𝘁𝗲𝗱 𝗮𝗹𝗼𝗻𝗴𝘀𝗶𝗱𝗲 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆. 🔗 Mistral Large 4 release 👉 https://lnkd.in/ekKttJzt 🔗 Benchmark 👉 https://b3.lakera.ai/ #AISecurity #AgenticAI #LLMSecurity #PromptInjection #GenAI

Mistral reporting b3 as a core security benchmark for the release of Mistral 4, exciting times!

An AI application that only answers questions has a fairly contained job. But once you connect it to the CRM, the ticketing queue, a payments API and a couple of internal workflows, it starts reading, changing and sending things on its own. The security questions change along with it: what it can reach, who decided that, and what happens when a run of ordinary actions adds up to something nobody intended. Those questions are where our next webinar, Securing AI Apps & Agents, starts. Steve Giguere, Principal AI Security Advocate, and Sam Watts, Product Lead for AI Agent Security, walk through what it takes to see, govern and protect AI applications and agents once they are wired into enterprise data and tools, and how runtime security can stop risky agent behavior as it unfolds. 🔗 𝗢𝗻𝗰𝗲 𝗶𝘁 𝗰𝗮𝗻 𝗮𝗰𝘁, 𝘄𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀 𝗶𝘀 𝘁𝗵𝗲 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻. 🗓️ Wednesday, October 14 · 10 AM PT / 1 PM ET / 7 PM CEST · 40 minutes, live. Registration link in the first comment. 👇 #AISecurity #AgenticAI #CyberSecurity

𝗪𝗵𝗮𝘁 𝗶𝗳 𝗮𝗻 𝗶𝗺𝗮𝗴𝗲 𝗰𝗼𝘂𝗹𝗱 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁 𝘆𝗼𝘂𝗿 𝗔𝗜 𝗮𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁? During our recent Check Point AI Security hackathon, one team explored whether a prompt could be hidden inside an ordinary image. The image looked completely normal. After the image was resized for the vision model, hidden text became readable to the assistant. When asked only to describe the picture, the assistant produced a normal response and quietly created a new calendar event containing the user's schedule. ⚠️ 𝗔 𝗳𝗲𝘄 𝗽𝗶𝘅𝗲𝗹𝘀 𝘄𝗲𝗿𝗲 𝗲𝗻𝗼𝘂𝗴𝗵 𝘁𝗼 𝗶𝗻𝗳𝗹𝘂𝗲𝗻𝗰𝗲 𝘁𝗵𝗲 𝗮𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁'𝘀 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝘂𝗿. The assistant still acted on it. A lot of discussion around AI security focuses on prompts. This experiment started with pixels. That was one of many directions explored during the hackathon. Other teams examined different ways AI systems can be influenced, manipulated or pushed beyond the assumptions their operators started with. There were plenty of other surprises. We collected some of them in the blog post below. 🔗 𝗟𝗶𝗻𝗸 𝗶𝗻 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝗰𝗼𝗺𝗺𝗲𝗻𝘁 👇 #AISecurity #AgentSecurity #PromptInjection #CyberSecurity #AIAgents

No single layer can watch an agent from every angle, and no agent can be trusted to watch itself. That's the idea behind our expanded work with NVIDIA on the Open Agent Safety Platform: independent controls for AI agents, and infrastructure-level protection for the AI factories behind them. NVIDIA OpenShell secures the agent layer, NVIDIA BlueField secures the infrastructure, and Check Point brings behavior monitoring and AI Factory security to both. Proud to build the agentic era's security architecture with NVIDIA. 🤝 Full announcement in the comments below 👇

Clalit cares for five million people, and like healthcare workers everywhere, its doctors and staff are already turning to public AI tools, which is exactly where sensitive patient data can slip out without anyone meaning it to. Add AI agents and third-party apps acting on their own, and the question for a hospital security team stops being whether to allow AI and becomes how to let it in safely. That is what Clalit, Israel's largest health fund, and Check Point are now building together - a security gateway that monitors AI use in real time, stops unauthorized data from leaving, sets clear usage policies and blocks anomalous actions by external AI agents. As Clalit CISO Avivit Kotler points out, Check Point sees only metadata, never patient data. 🔒 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 𝘁𝗵𝗮𝘁 𝗹𝗲𝘁 𝗰𝗹𝗶𝗻𝗶𝗰𝗶𝗮𝗻𝘀 𝘂𝘀𝗲 𝗔𝗜 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝘁𝗿𝗮𝗱𝗶𝗻𝗴 𝗮𝘄𝗮𝘆 𝗽𝗮𝘁𝗶𝗲𝗻𝘁 𝗽𝗿𝗶𝘃𝗮𝗰𝘆. The full story in Calcalist Tech 👉 https://lnkd.in/duHJX9ru #AISecurity #Healthcare #CyberSecurity #DataPrivacy #HealthTech

An agent can do real damage while doing its best, and one of our hackathon teams has the traces to show it. They gave an autonomous SRE agent tasks it could not possibly complete and let it work in a sandbox built so that dangerous actions could play out safely. Nobody attacked it and nobody tried to trick it. It rebooted a production database, dropped customer connections and escalated its own privileges, at one point guessing its way through roles to find the access it wanted. At every step, its reasoning read like a diligent engineer talking itself into the next move. That is what makes this kind of failure so hard to catch. The intent stayed good the whole time, so the thing worth judging is the action itself, and whether it fits the task the agent was actually given. 🔌 𝗚𝗼𝗼𝗱 𝗶𝗻𝘁𝗲𝗻𝘁𝗶𝗼𝗻𝘀 𝗮𝗿𝗲 𝗻𝗼𝘁 𝗮 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗰𝗼𝗻𝘁𝗿𝗼𝗹. It is one of three findings from our August hackathon, all in the write-up linked in the first comment. 👇 #AISecurity #AgenticAI #CyberSecurity