ActiveFence

👉 GenAI Friday Roundup | October 17 Friday’s here! Time to catch up on the latest in AI news, launches, and policy moves. Stay informed, stay ahead. 1️⃣ California’s AI law package sets new standards for safety and transparency - Governor Gavin Newsom signed SB53, requiring large AI firms to disclose catastrophic risk mitigation plans, while also advancing rules on chatbot content with minors and origin labeling with SB 243. Together, these moves signal a pragmatic turn in state-level AI governance: pairing ambition with enforceability and driving demand for compliance tooling and audit-ready safety processes. 🔗https://lnkd.in/gN_eTVVi https://lnkd.in/gU7ybcSx 2️⃣ White House and Anthropic diverge on AI oversight - A growing rift between the White House and safety-first labs like Anthropic underscores a policy crossroads: centralized regulation vs. industry-led governance. The outcome could define how compliance, accountability, and safety baselines evolve in 2025. 🔗https://lnkd.in/g2-f4FCf 3️⃣ Organizations struggle with agentic AI governance - A new survey shows 59% of business leaders lack policies for agentic AI, while 87% of employees don’t know existing rules. Proactive AI safety frameworks become more essential as enterprises embrace autonomous tools, and governance gaps widen.  🔗https://lnkd.in/gK4yM8Hs 4️⃣ Hidden prompts turn Perplexity’s AI browser into a phishing tool - ActiveFence researchers were able to inject prompts in Perplexity’s Comet AI browser that could be exploited to generate realistic phishing content and harvest personal data. ActiveFence is now working closely with Perplexity to ensure the issue is fully resolved. It’s a reminder that even user-facing AI assistants can be manipulated into social engineering vectors, showing why real-time guardrails and adversarial prompt detection matter so much. 🔗 https://lnkd.in/gSFS-FhM 5️⃣ UK spy chief warns of AI threat - MI5 chief Ken McCallum cautioned that while AI isn’t heading toward Terminator scenarios, autonomous systems already pose real risks in propaganda, surveillance, and election interference.  For enterprises and governments alike, ensuring information and data integrity is a core part of AI risk management. 🔗https://lnkd.in/gVAapKEF 💡 Our take: From regulators to courts to enterprises, AI oversight is shifting from abstract principles to concrete practice. The new frontier of AI safety is about smarter models and smarter governance: attribution, accountability, and assurance for every interaction, at every layer.

𝗬𝗼𝘂 𝗻𝗲𝗲𝗱 𝗮 𝗰𝗹𝗲𝗮𝗿 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝘁𝗼 𝗺𝗮𝗻𝗮𝗴𝗲 𝘀𝗲𝗻𝘀𝗶𝘁𝗶𝘃𝗲 𝗱𝗮𝘁𝗮 𝗮𝗻𝗱 𝗳𝗮𝘀𝘁-𝗰𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝘁𝗵𝗿𝗲𝗮𝘁𝘀 as your team builds more customer-facing AI. An AI Safety and Security policy gives you that. 𝗕𝘂𝘁 𝘄𝗵𝗲𝗿𝗲 𝗱𝗼 𝘆𝗼𝘂 𝘀𝘁𝗮𝗿𝘁? 𝗪𝗵𝗲𝗿𝗲 𝗰𝗮𝗻 𝘆𝗼𝘂 𝗶𝗺𝗽𝗿𝗼𝘃𝗲? Start by asking questions on how your organization builds, tests, and governs AI systems. Look at where data comes from, who has access, how outputs are monitored, and what safeguards exist when things go wrong. The ActiveFence AI Safety and Security Questionnaire walks teams through these foundations, helping leaders identify blind spots and set stronger policies. Use it as a checklist to build an effective, dynamic policy and align teams on data, red teaming, and incident response. Get it here → https://lnkd.in/g-VxGTis

𝗟𝗲𝘁’𝘀 𝘁𝗮𝗹𝗸 𝗮𝗹𝗹 𝘁𝗵𝗶𝗻𝗴𝘀 𝗔𝗚𝗘𝗡𝗧𝗜𝗖: 𝘀𝘆𝘀𝘁𝗲𝗺𝘀, 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀, 𝗮𝗻𝗱 𝗯𝗿𝗼𝘄𝘀𝗲𝗿𝘀. We’re heading to AGENTIC 2025, where the brightest minds in AI are bridging the gap between autonomy and accountability. Join John O'Donnell and Tomer Poran as we explore what’s next for enterprise GenAI. 📍Where: 𝗔𝗿𝗹𝗶𝗻𝗴𝘁𝗼𝗻, 𝗩𝗔 📆 When: 𝗢𝗰𝘁𝗼𝗯𝗲𝗿 𝟮𝟳–𝟮𝟵! 👉🏻 Learn more - and book a meeting: https://lnkd.in/gBrE6PJu #GoToAGENTIC #AgenticAI #GenAI #AItrust #AIsecurity #ActiveFence

Most prompt engineering stops once a prompt works. But the real gains come when prompts keep learning and improving. See how our process for feedback-driven prompt optimization boosts accuracy, without any code changes. Learn more in our on-demand webinar, hosted by Shiri Simon Segal. 👉 https://lnkd.in/gedmrq29 #GenAI #MachineLearning #PromptDesign #AIInnovation

👉 𝗚𝗲𝗻𝗔𝗜 𝗙𝗿𝗶𝗱𝗮𝘆 𝗥𝗼𝘂𝗻𝗱𝘂𝗽 | 𝗢𝗰𝘁𝗼𝗯𝗲𝗿 𝟭𝟬 It's Friday! Catch up on the latest AI news, product releases, and policy changes before the weekend, so you can stay one prompt ahead of the AI curve. 1️⃣ 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗯𝗲𝗰𝗼𝗺𝗲𝘀 𝘁𝗵𝗲 𝗻𝗲𝘄 𝗢𝗦 – At DevDay, OpenAI unveiled an app store and in-chat apps for #ChatGPT, creating new functionalities and opening up new attack surfaces. Platform safety and runtime governance will now matter as much as model safety. https://lnkd.in/gYeZ2qh3 2️⃣ 𝗖𝗼𝗱𝗲𝘅 𝗚𝗔 𝗮𝗻𝗱 𝗔𝗴𝗲𝗻𝘁𝗞𝗶𝘁 𝗹𝗮𝘂𝗻𝗰𝗵 – Also at DevDay, OpenAI rolled out tools to make #Agentic AI accessible to developers everywhere. With broader use comes broader risk exposure, like prompt chaining and unsupervised actions. https://lnkd.in/gzPbgs_R 3️⃣ Deloitte 𝘁𝗼 𝗿𝗲𝗳𝘂𝗻𝗱 𝗴𝗼𝘃𝗲𝗿𝗻𝗺𝗲𝗻𝘁 𝗮𝗳𝘁𝗲𝗿 𝗔𝗜-𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝗿𝗲𝗽𝗼𝗿𝘁 𝗲𝗿𝗿𝗼𝗿𝘀 – A reminder that even top consultancies can fall victim to AI hallucinations. Enterprises need systems for validation, provenance, and human review before trusting generative outputs. https://lnkd.in/gVSvBThr 4️⃣ 𝗦𝘁𝗮𝘁𝗲-𝗹𝗲𝘃𝗲𝗹 𝗔𝗜 𝗹𝗮𝘄𝘀 𝗽𝗿𝗼𝗹𝗶𝗳𝗲𝗿𝗮𝘁𝗲 – U.S. states are rolling out their own AI governance frameworks, from liability rules to automated decision audits. Compliance and accountability are becoming table stakes for responsible deployment. https://lnkd.in/gkdB57ns 💡 𝗢𝘂𝗿 𝘁𝗮𝗸𝗲: As AI becomes infrastructure, powering apps, workflows, and governance itself, safety and accountability must scale in parallel. Guardrails can’t just sit around the model; they have to run through every layer: platforms, agents, enterprises, and regulators alike.

𝗙𝗼𝗿 𝗱𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹𝘀, 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝗶𝘀 𝗷𝘂𝘀𝘁 𝗼𝗻𝗲 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗲𝗾𝘂𝗮𝘁𝗶𝗼𝗻. They also need to be fast, scalable, and cost-efficient to perform in real-time environments. Learn from Shiri Simon Segal how her team is tackling this challenge in our on-demand webinar, 𝗗𝗶𝘀𝘁𝗶𝗹𝗹𝗶𝗻𝗴 𝗟𝗟𝗠𝘀 𝗶𝗻𝘁𝗼 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿𝘀 𝗳𝗼𝗿 𝗥𝗲𝗮𝗹-𝗪𝗼𝗿𝗹𝗱 𝗔𝗜. 𝗪𝗮𝘁𝗰𝗵 𝗶𝘁 𝗻𝗼𝘄: https://lnkd.in/gedmrq29 #GenAI #AIOptimization #MachineLearning #ActiveFence

𝗪𝗲'𝘃𝗲 𝘀𝗲𝗲𝗻 𝘁𝗵𝗮𝘁 𝗔𝗜 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝘀 𝗰𝗮𝗻 𝗳𝗮𝗶𝗹 𝗶𝗻 𝘂𝗻𝗲𝘅𝗽𝗲𝗰𝘁𝗲𝗱 𝘄𝗮𝘆𝘀. Jailbreaks, hidden misuse, prompt injection—these are just some of the risks associated with launching an AI application without proper testing. And, as AI becomes embedded in users' everyday lives, minor gaps can turn into huge risks - fast. #RedTeaming is a critical tool to spot these risks, before they become a liability. Download our report and learn core red teaming tools and processes, straight from our expert team: https://lnkd.in/gZRync3U #GenAI #AISafety #ResponsibleAI

Coming to World Summit AI tomorrow, or in the Amsterdam area? Get a discounted pass with our code (below), swing by our booth, and let's talk #guardrails! Schedule a meeting here: https://lnkd.in/gwVW3r-f

🚨 Perplexity’𝘀 𝗖𝗼𝗺𝗲𝘁 𝗯𝗿𝗼𝘄𝘀𝗲𝗿 𝗷𝘂𝘀𝘁 𝗿𝗼𝗹𝗹𝗲𝗱 𝗼𝘂𝘁 𝗳𝗿𝗲𝗲 𝘁𝗼 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝗹𝗮𝘀𝘁 𝘄𝗲𝗲𝗸, putting AI-powered browsing into the hands of millions. That’s exciting… but also risky. Our research shows how Comet’s trust loop can be exploited in the free version: 𝗶𝗻𝘃𝗶𝘀𝗶𝗯𝗹𝗲 𝗽𝗮𝘆𝗹𝗼𝗮𝗱𝘀, 𝗽𝗵𝗶𝘀𝗵𝗶𝗻𝗴 𝗹𝗶𝗻𝗸𝘀, 𝗮𝗻𝗱 𝗲𝘃𝗲𝗻 𝗵𝗶𝗱𝗱𝗲𝗻 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻𝘀 inside working documents. Check out our attack flow below, and read the full breakdown to see how misplaced trust in AI browsers can become a real threat. https://lnkd.in/gzhqy-ms #AI #Security #AgenticAI #BrowserSecurity

🌎 𝗢𝗰𝘁𝗼𝗯𝗲𝗿 𝗶𝘀 𝗽𝗮𝗰𝗸𝗲𝗱. From Amsterdam to Florida, our team will be on stage, at the booth, and in the room for the season’s biggest AI & security gatherings. 👉 Come say hi, swap ideas, or just grab a coffee with us.   Check out where we’ll be: https://lnkd.in/dfts5J6C #GenAI #AISafety #EnterpriseAI #AIevents #ActiveFenceOnTheRoad