𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗱𝗮𝘁𝗮 𝗹𝗲𝗮𝗸𝗮𝗴𝗲 𝗹𝗼𝗼𝗸 𝗹𝗶𝗸𝗲 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂𝗿 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝘁𝗵𝗶𝗻𝗸 𝗶𝗻 𝗻𝗮𝘁𝘂𝗿𝗮𝗹 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲? Not like it used to. Sensitive data doesn’t just sit in files or emails anymore. It flows through model outputs: summaries, translations, and multi-step agent workflows. It gets paraphrased, transformed, and leaked, long before traditional DLP even has a chance to react. We just published a new piece on what modern DLP should look like for GenAI systems: ✅ Why traditional tools fall short ✅ How leakage happens through language and reasoning ✅ A practical checklist for securing GenAI apps ✅ What language-native, real-time DLP really means 📽️ The video below is a sneak preview of what’s coming to Lakera Guard: real-time, 𝗟𝗟𝗠-𝗽𝗼𝘄𝗲𝗿𝗲𝗱 𝗰𝘂𝘀𝘁𝗼𝗺 𝗱𝗲𝘁𝗲𝗰𝘁𝗼𝗿𝘀 you can configure in plain English. No rulesets. No training data. Just natural descriptions and instant protection. 🔗 Link to the full article in the first comment.
Lakera
Software Development
Customers rely on Lakera for real-time security that doesn’t slow down their GenAI applications.
About us
Lakera is the world’s leading real-time GenAI security company. Customers rely on the Lakera AI Security Platform for security that doesn’t slow down their AI applications. To accelerate secure adoption of AI, the company created Gandalf, an educational platform, where more than one million users have learned about AI security. Lakera uses AI to continuously evolve defenses, so customers can stay ahead of emerging threats. Join us to shape the future of intelligent computing: www.lakera.ai/careers
- Website
-
https://lakera.ai
External link for Lakera
- Industry
- Software Development
- Company size
- 11-50 employees
- Headquarters
- San Francisco
- Type
- Privately Held
- Founded
- 2021
- Specialties
- llm, GenAI, AI security, machine learning, and artificial intelligence
Locations
-
Primary
Get directions
San Francisco, US
Employees at Lakera
Updates
-
Lakera reposted this
3, 2, 1...and it’s going to be a big week at RSAC! 🎥 I look forward to meeting with friends, colleagues, and partners while I’m there. Here is where you can find me: 🌮 𝗦𝗲𝗰𝘂𝗿𝗲 𝘁𝗵𝗲 𝗦𝗮𝗹𝘀𝗮: 𝗔𝗜 𝘅 𝗖𝘆𝗯𝗲𝗿 𝗛𝗮𝗽𝗽𝘆 𝗛𝗼𝘂𝗿 – co-hosted with Fly Ventures 📅 Monday, April 28 | 6:00pm - 8:00pm PDT 📍 Lakera HQ (5’ from Moscone) Join AI builders & investors for tacos, guacamole, and beers. Let me know if you’d like to join. 🍽️ 𝗔𝗜 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗟𝗲𝗮𝗱𝗲𝗿𝘀' 𝗟𝘂𝗻𝗰𝗵 – co-hosted with Joe Sullivan 📅 Wednesday, April 30 | 12:00pm - 2:00pm PDT 📍 Lakera HQ (5’ from Moscone) Lakera's invite-only lunch with security leaders that are building and securing the future of agents AI. Really excited about this one. A few spots remaining, send me a DM if you are interested in joining. 🔟 𝗢𝗪𝗔𝗦𝗣 𝗔𝗜 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗦𝘂𝗺𝗺𝗶𝘁 – hosted by OWASP GenAI Security Project 📅 Wednesday, April 30 | 8:30 AM - 12:30 PM PDT 📍 Moscone South, Room #360 Dive deep and explore the latest best practices in LLM & GenAI security, agentic security, AI red teaming, incident response, and governance. With Scott Clinton, Steve Wilson, and Ads Dawson 🛥️ 𝗔𝗜 𝗖𝘆𝗯𝗲𝗿 𝗬𝗮𝗰𝗵𝘁 – co-hosted with Sergej Epp 📅 Wednesday, April 30 | 6:00 PM - 9:00 PM PDT 📍 Pier 40 Looking forward to an unforgettable night with AI and security leaders cruising the San Francisco Bay. A few spots remaining, send me a DM if you want to join. If you’re attending RSAC, let’s connect. Drop me a note or come by one of our sessions. #RSAC2025 #GenAI #AISecurity #OWASP
-
-
🔐 𝗪𝗵𝘆 𝗬𝗼𝘂𝗿 𝗔𝗜 𝗗𝗲𝗳𝗲𝗻𝘀𝗲𝘀 𝗔𝗿𝗲 𝗢𝘂𝘁𝗱𝗮𝘁𝗲𝗱 𝘁𝗵𝗲 𝗠𝗼𝗺𝗲𝗻𝘁 𝗬𝗼𝘂 𝗗𝗲𝗽𝗹𝗼𝘆 𝗧𝗵𝗲𝗺 GenAI models change all the time—so do the threats. But here’s the problem: most security strategies assume a static world. Traditional defenses rely on: ❌ 𝗕𝗹𝗼𝗰𝗸𝗹𝗶𝘀𝘁𝘀 – Attackers just rephrase their prompts. ❌ 𝗥𝘂𝗹𝗲-𝗯𝗮𝘀𝗲𝗱 𝗱𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 – New model updates introduce unforeseen vulnerabilities. ❌ 𝗙𝗶𝘅𝗲𝗱 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝘀𝗲𝘁𝘁𝗶𝗻𝗴𝘀 – But AI doesn’t work that way—it learns, shifts, evolves. So how do you defend a system that evolves by the day? 𝗪𝗲 𝗱𝗼𝗻’𝘁 𝗷𝘂𝘀𝘁 𝗻𝗲𝗲𝗱 𝗯𝗲𝘁𝘁𝗲𝗿 𝗱𝗲𝗳𝗲𝗻𝘀𝗲𝘀—𝘄𝗲 𝗻𝗲𝗲𝗱 𝗮𝗱𝗮𝗽𝘁𝗶𝘃𝗲 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆. 💡 That’s why at Lakera, we use red teaming to simulate real-world attacks. Instead of chasing vulnerabilities after they’ve been exploited, we identify them before they happen. ⏭️ 𝗧𝗵𝗶𝘀 𝗶𝘀 𝗣𝗼𝘀𝘁 𝟮 𝗼𝗳 𝟱 in our 𝘈𝘐 𝘙𝘦𝘥 𝘛𝘦𝘢𝘮𝘪𝘯𝘨 𝘚𝘦𝘳𝘪𝘦𝘴. 𝗡𝗲𝘅𝘁: What real-world AI Red Teaming actually looks like—and how attackers are already exploiting AI systems today. 🎥 In this clip, Lakera's CEO David Haber explains how Gandalf reveals vulnerabilities in real time. 📌 The first comment has a guide with lessons from Gandalf—and insights from the world’s largest AI red team.
-
LLaMA 4 is fast. It’s smart. But is it secure? Mateo Rojasjust shared some early insights from the benchmark we’ve been building at Lakera Research, and the results are raising serious questions about the tradeoffs between model performance and real-world security. We’ll be publishing the full benchmark soon. 📩 Want to be the first to know when it drops? Shoot us a message and we’ll make sure you don’t miss it.
🦙 LLaMA 4 brings decisive evidence: newer, more powerful LLMs are not always more secure. This challenges a quiet but widespread assumption in our field: that security naturally improves with model performance. LLaMA 4 just dropped, and as expected, the quality benchmarks followed. Some in the community are already raising good questions about the validity of those numbers. But even if they hold up, those benchmarks measure just one thing: performance. At Lakera Research, we’ve been thinking about something else entirely: security. Current benchmarks barely scratch the surface when it comes to security. So we built our own. Our benchmark measures how successful an attacker is across a wide range of goals—exfiltrating secrets, bypassing safety systems, gaining control over the model, and more. More details on our benchmark will follow soon, but here’s the early insight we need to talk about: LLaMA 4 is markedly less secure than its LLaMA 3 predecessor. That’s a big deal. It shows that optimism around "security through model progress" is far from guaranteed. In fact, newer models might introduce entirely new attack surfaces. Models are getting faster, smarter, and more capable. But not more secure. It’s time we expand how we define progress in AI.
-
-
🚨 𝗧𝗿𝗮𝗱𝗶𝘁𝗶𝗼𝗻𝗮𝗹 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗪𝗼𝗻’𝘁 𝗦𝗮𝘃𝗲 𝗬𝗼𝘂𝗿 𝗚𝗲𝗻𝗔𝗜 𝗦𝘆𝘀𝘁𝗲𝗺. 𝗛𝗲𝗿𝗲’𝘀 𝗪𝗵𝘆. For years, companies have relied on 𝗽𝗲𝗻𝗲𝘁𝗿𝗮𝘁𝗶𝗼𝗻 𝘁𝗲𝘀𝘁𝗶𝗻𝗴 to find security flaws. But what happens when the biggest risk isn’t in your code—but in how people interact with your AI? 𝗚𝗲𝗻𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗮𝗿𝗲 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 because attackers don’t need to break in. They just need to talk their way in. Every prompt a user enters is 𝗮𝗻 𝗼𝗽𝗽𝗼𝗿𝘁𝘂𝗻𝗶𝘁𝘆 𝘁𝗼 𝗼𝘃𝗲𝗿𝗿𝗶𝗱𝗲 𝘆𝗼𝘂𝗿 𝗔𝗜’𝘀 𝗶𝗻𝘁𝗲𝗻𝗱𝗲𝗱 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝗿. This means: 🔹 Every user is a potential hacker 🔹 Every input is a possible attack 🔹 Traditional defenses can’t keep up So how do we defend against a 𝘁𝗵𝗿𝗲𝗮𝘁 𝗹𝗮𝗻𝗱𝘀𝗰𝗮𝗽𝗲 𝘁𝗵𝗮𝘁 𝗰𝗼𝗻𝘀𝘁𝗮𝗻𝘁𝗹𝘆 𝗲𝘃𝗼𝗹𝘃𝗲𝘀? ⏭️ This is Post 1 of 5 in our AI Red Teaming Series. 𝗨𝗽 𝗻𝗲𝘅𝘁: 𝗪𝗵𝘆 𝘀𝘁𝗮𝘁𝗶𝗰 𝗱𝗲𝗳𝗲𝗻𝘀𝗲𝘀 𝗳𝗮𝗶𝗹—𝗮𝗻𝗱 𝘄𝗵𝗮𝘁 𝘄𝗲 𝗻𝗲𝗲𝗱 𝗶𝗻𝘀𝘁𝗲𝗮𝗱. Matt F. explains why every prompt is like committing code to your GenAI system—watch the clip below. Curious how Lakera Red tests your AI’s resilience? Link’s in the first comment.👇
-
⏳ The countdown to RSAC is on — just one month to go. We’re heading back to San Francisco for 𝗥𝗦𝗔𝗖 𝟮𝟬𝟮𝟱, and we’re bringing the full force of Lakera with us. 📍 𝗕𝗼𝗼𝘁𝗵 𝗦-𝟮𝟰𝟱𝟯 Here’s what you can expect: 🔹 A behind-the-scenes look at 𝗟𝗮𝗸𝗲𝗿𝗮 𝗚𝘂𝗮𝗿𝗱 and 𝗟𝗮𝗸𝗲𝗿𝗮 𝗥𝗲𝗱 🔹 Live demos of real GenAI attacks — and how to stop them 🔹 Conversations with our AI security experts 🔹 A chance to 𝗷𝗼𝗶𝗻 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱’𝘀 𝗹𝗮𝗿𝗴𝗲𝘀𝘁 𝗔𝗜 𝗿𝗲𝗱 𝘁𝗲𝗮𝗺 Whether you’re building, breaking, or securing GenAI systems — we’d love to chat. 👉 Book a demo or check out what we’re up to: https://lnkd.in/dbZV9_mp See you at Moscone Center. #RSAC2025 #CyberSecurity #AIsecurity #GenAI #AI #RedTeam #ThreatDetection #RSAC #CyberDefense #Lakera #AIThreats #Infosec #SecurityInnovation #MachineLearning
-
-
Big shoutout to Sam Watts for bringing #Gandalf to the party — and showing a room full of AI experts how easily things can break. You shall not pass… unless you’re really good at prompt injections. 😈
YOU SHALL NOT PASS!!! 🤯 A room full of brilliant minds — but the big message hit home: We’ve opened the door to infinite attack vectors. 🎁 Now what? 💉Prompt injection isn’t just a theoretical problem — it’s a full-blown art form. At 'AI Security Unplugged: Agentic Models & Red Teaming' we saw everything from obfuscation to roleplay attacks charted in glorious technicolor. Spyros Gasteratos took us through the work he's doing at Smithy and practical advice on how to protect yourself and users from the inherent vulnerabilities that exist with LLMs (check out his blog here: https://lnkd.in/ePNn8X8m). And Sam Watts introduced me to Gandalf🧙♂️ (no relation...), the platform to test out your jailbreaking skills, give it a go: https://lnkd.in/euWe_pNZ The most popular strategies: Just ask the AI nicely...or in French....or to act as <insert author name> and craft a magical story that includes the "special sauce". Kudos to AI Security Collective, Investigo and ControlPlane for bringing the heat (and the data). #PromptHacking #AIredteam #CyberThreats #AIvulnerability #CyberSecurity #AICommunity #LLM
-
-
If you’re working with RAG and haven’t seen this yet—now’s the time. 👇 Real-world attacks, real consequences. This is one of the clearest breakdowns of how the LLM vulnerability plays out in production systems. Highly recommended read from the team:
🛡️ LLM security is a hot topic, but it’s not always clear how these models are actually vulnerable in real-world applications. In the second post of our blog series on building a superhuman red teaming agent, Peter Dienes gives a great overview of real, tangible risks that affect RAG applications. To me, the highlight is that these systems are especially vulnerable to data from third-party sources, data that developers don’t control, but attackers can exploit. Check out the full post here: 👉 https://lnkd.in/eEbT7bdf Next, we’ll explore how to automatically evaluate LLM red teamers—and why building strong automated attackers is so challenging. Stay tuned, and let us know what you think!
-
-
𝘞𝘩𝘢𝘵 𝘥𝘰 𝘺𝘰𝘶 𝘤𝘢𝘭𝘭 𝘢𝘯 𝘢𝘵𝘵𝘢𝘤𝘬 𝘵𝘩𝘢𝘵 𝘩𝘪𝘥𝘦𝘴 𝘮𝘢𝘭𝘪𝘤𝘪𝘰𝘶𝘴 𝘪𝘯𝘵𝘦𝘯𝘵 𝘪𝘯𝘴𝘪𝘥𝘦 𝘢 𝘤𝘰𝘯𝘧𝘪𝘨 𝘧𝘪𝘭𝘦—𝘢𝘯𝘥 𝘵𝘶𝘳𝘯𝘴 𝘺𝘰𝘶𝘳 𝘈𝘐 𝘢𝘴𝘴𝘪𝘴𝘵𝘢𝘯𝘵 𝘪𝘯𝘵𝘰 𝘢 𝘣𝘢𝘤𝘬𝘥𝘰𝘰𝘳? We’ve seen this coming for a while. A new vulnerability in GitHub Copilot and Cursor shows how easy it is to exploit AI systems via the 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻𝘀 𝘁𝗵𝗲𝘆’𝗿𝗲 𝗳𝗲𝗱—𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝘁𝗵𝗲 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 𝘂𝘀𝗲𝗿𝘀 𝘁𝘆𝗽𝗲. 🔍 𝗧𝗵𝗲 𝗮𝘁𝘁𝗮𝗰𝗸: Hackers inject malicious logic into 𝗥𝘂𝗹𝗲𝘀 𝗙𝗶𝗹𝗲𝘀—configuration files that guide how code agents like Copilot and Cursor behave. By embedding invisible unicode and remote script loaders, attackers can manipulate the AI into generating malicious code… without ever touching the user prompt. It’s not a typical prompt injection. It’s a 𝗺𝗼𝗱𝗲𝗹-𝗳𝗮𝗰𝗶𝗻𝗴 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻 𝗮𝘁𝘁𝗮𝗰𝗸—and it bypasses most traditional safeguards. These files are: 🔹 Shared across teams 🔹 Open-sourced and implicitly trusted 🔹 Rarely reviewed through a security lens …which makes them the perfect place to hide. 💣 This isn’t hypothetical. It’s the new reality of AI development. 𝗔𝗻𝗱 𝗶𝘁’𝘀 𝗲𝘅𝗮𝗰𝘁𝗹𝘆 𝘄𝗵𝘆 𝘄𝗲 𝗯𝘂𝗶𝗹𝘁 𝗟𝗮𝗸𝗲𝗿𝗮 𝗚𝘂𝗮𝗿𝗱: ✅ To analyze instruction payloads, not just prompts ✅ To defend against threats embedded in system context ✅ To catch what static code scanning never will 👇 How would you classify this type of attack? Let’s discuss—because while the threat is evolving, the defenses are already here. Read the full article here 👉 https://lnkd.in/gzSmmJZS
-
🚀 𝗜𝘁’𝘀 𝗵𝗲𝗿𝗲: 𝗠𝘂𝗹𝘁𝗶𝗹𝗶𝗻𝗴𝘂𝗮𝗹 𝗔𝗜 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗶𝘀 𝗟𝗜𝗩𝗘! 🌍🔒 Last week, we told you to keep an eye out for something big. Well, today’s the day. 𝗟𝗮𝗸𝗲𝗿𝗮 𝗚𝘂𝗮𝗿𝗱 𝗻𝗼𝘄 𝘀𝗲𝗰𝘂𝗿𝗲𝘀 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 𝗮𝗴𝗮𝗶𝗻𝘀𝘁 𝗽𝗿𝗼𝗺𝗽𝘁 𝗮𝘁𝘁𝗮𝗰𝗸𝘀 𝗶𝗻 𝟭𝟬𝟬+ 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲𝘀! 🎉 That means: ✅ 𝗚𝗹𝗼𝗯𝗮𝗹 𝗔𝗜 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆—whether your GenAI is used in English, Japanese, Arabic, Hindi, or Swahili, it’s now protected. ✅ 𝗦𝘁𝗿𝗼𝗻𝗴𝗲𝗿 𝗱𝗲𝗳𝗲𝗻𝘀𝗲𝘀—we’ve improved detection accuracy while reducing false positives for a smoother experience. ✅ 𝗔𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 𝗳𝗼𝗿 𝗮𝗹𝗹 𝘂𝘀𝗲𝗿𝘀—if you’re a free community user, evaluating Lakera Guard, or an enterprise customer, you’ve been updated. No action needed! This update is a huge step forward in securing AI systems at a global scale. And we’re just getting started—stay tuned for even more updates on custom guardrails and more! Ready to test your AI defenses? Try it now. And let us know—what language should we challenge your AI in first? 👇 #AIsecurity #GenAI #MultilingualSecurity #LakeraGuard
翻译：