Really enjoyed your Friday night talk @UNB, Dr.@lemire. Also ordered Rainbows End!

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥 AI needs to learn like humans—or we lose the race. If you can read a book, learn from it, and use that knowledge in your work, why shouldn’t AI be able to do the same? @sama and @OpenAI are making the case that restricting AI training data is a direct threat to innovation and national security. They’re right. AI models don’t hoard copyrighted works; they learn from them—just like humans. The result? A lossy, mathematical representation of concepts, not a direct copy-paste. The real issue isn’t whether AI should be able to train on data—𝗶𝘁’𝘀 𝗲𝗻𝘀𝘂𝗿𝗶𝗻𝗴 𝗳𝗮𝗶𝗿 𝘂𝘀𝗲 𝗽𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲𝘀 𝗮𝗿𝗲 𝗮𝗽𝗽𝗹𝗶𝗲𝗱 𝗰𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝘁𝗹𝘆. AI should have access to all publicly available data—just like humans—except for private or pirated content, which wouldn't be fair use for people either. At the same time, AI should be held to a similar standard, i.e., no infringement of copyright, including derivative content, and crediting sources appropriately. But while the U.S. debates, China isn’t waiting. If we limit AI’s ability to learn while our adversaries train on everything, we’re setting ourselves up to lose the AI race—and that’s a serious national security risk. The solution? 𝗥𝗲𝗴𝘂𝗹𝗮𝘁𝗲 𝘁𝗵𝗲 𝗼𝘂𝘁𝗽𝘂𝘁, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗽𝗿𝗼𝗰𝗲𝘀𝘀. Hold it accountable for what it produces, not for what it reads. Otherwise, we’re kneecapping AI innovation in a world where intelligence wins. Let’s get this right. Follow us over @TrojAISec for more. #AI #FairUse #Cybersecurity #AIEthics #OpenAI

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥 Prompt Injection: The Threat You Can’t Triage Away I saw an AI security influencer downplaying prompt injection recently. I get that CISOs have a mountain of threats to triage and that prioritization is key to survival but outright dismissing this risk is 𝘥𝘢𝘯𝘨𝘦𝘳𝘰𝘶𝘴. Prompt injection isn’t some fringe security issue. It’s the defining security challenge of GenAI. @owasp—who has shaped AppSec for decades—defines prompt injection as the number one most critical AI risk since first releasing their Top 10 for LLMs list. Here’s the reality: As AI systems get more complex, the attack surface expands. We’re no longer just talking about chatbots getting tricked into saying something dumb. We’re talking about supply chain attacks on AI-driven automation, financial fraud via LLM-powered workflows, and the ability to manipulate critical decision-making systems. If you don’t yet see prompt injection as a major issue, take another look. Prompt injection isn’t an isolated vulnerability. It’s a fundamental flaw in how LLMs process input. The more AI integrates into business logic, the harder it will be to contain these attacks. I respect influencers—they help spread awareness. But cybersecurity isn’t just about hype cycles. It’s about knowing which threats you can and cannot afford to triage away. CISOs, if you’re listening: Prompt injection isn’t just another bullet on a risk register. It’s an architectural problem—one that requires immediate and decisive mitigations. Follow us over @TrojAISec for more hot takes!

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥 Trust is the True Currency of Cybersecurity In cybersecurity, trust isn’t a nice-to-have—it’s the difference between resilience and disaster. Without it, even the most advanced AI security solutions mean nothing. Customers don’t just buy protection; they invest in confidence that they can count on to avoid: 💥 𝗙𝗶𝗻𝗮𝗻𝗰𝗶𝗮𝗹 𝗹𝗼𝘀𝘀 – A single mistake can cost millions. ⚠️ 𝗥𝗲𝗽𝘂𝘁𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗱𝗮𝗺𝗮𝗴𝗲 – Customers don’t forgive easily when trust is broken. 🔎 𝗥𝗲𝗴𝘂𝗹𝗮𝘁𝗼𝗿𝘆 𝘀𝗰𝗿𝘂𝘁𝗶𝗻𝘆 – Compliance violations bring heavy fines and legal battles. That’s why we don’t just build security—we build trust. And it starts with our core values: 🔹 𝗧𝗲𝗮𝗺𝘄𝗼𝗿𝗸 – Because cybersecurity is a team sport. We stand together, stronger. 🔹 𝗖𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝗖𝗲𝗻𝘁𝗿𝗶𝗰𝗶𝘁𝘆 – We don’t chase trends; we solve real-world security challenges. 🔹 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 – The threat landscape evolves, so we evolve faster. 🔹 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻 – AI security isn’t 'traditional' security. It demands fresh, bold thinking. 🔹 𝗜𝗻𝘁𝗲𝗴𝗿𝗶𝘁𝘆 – If security isn’t built on honesty, it’s already broken. All of this, 𝘪𝘧 𝘭𝘪𝘷𝘦𝘥 𝘦𝘷𝘦𝘳𝘺𝘥𝘢𝘺, creates something our customers recognize instantly: authenticity. And that authenticity fuels trust. It’s why organizations choose us as their favorite cybersecurity partner—not just to protect data, but to secure the integrity of AI model behavior itself. Proud of this team. Proud of our mission. Join us over at TrojAI to learn how we are building trust in securing AI. #Cybersecurity #GenAI #CISO #OWASP #Infosec #HotTakeTuesdays

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥 Red Teaming AI: The Hype, The Reality, and What Actually Matters AI security is gaining momentum, and red teaming AI models is at the forefront of this shift. That’s great news. Protecting the integrity of model behavior is what makes AI security uniquely AI security, and we’re excited to see this focus growing across the industry. But as AI security takes center stage, it’s important to recognize that not all AI red teaming is the same. Red teaming is a discipline—built on deep expertise, creativity, and rigorous methodologies. AI is also a discipline—complex, evolving, and fundamentally different from traditional software. To effectively pentest AI systems, we need solutions that truly understand both. As more tools enter the market, security leaders have an opportunity to raise the bar. The best solutions will go beyond surface-level attacks and truly challenge AI models, uncovering vulnerabilities that impact real-world safety and reliability. Asking the right questions—𝗗𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝗱𝗲𝗲𝗽𝗹𝘆 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱 𝗔𝗜 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝗿? 𝗖𝗮𝗻 𝗶𝘁 𝗮𝗱𝗮𝗽𝘁 𝗹𝗶𝗸𝗲 𝗮 𝗿𝗲𝗮𝗹 𝗮𝗱𝘃𝗲𝗿𝘀𝗮𝗿𝘆?—helps cut through the noise and identify true best-in-class approaches. The future of AI security is being built now. With thoughtful evaluation and investment in true best-in-class methodologies, we can ensure AI remains secure, resilient, and trustworthy. Follow us over at @TrojAISec for more hot takes.

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥 Another week, another AI model caught with its guardrails down. Last week, everyone was talking about DeepSeek's new R1 model and its failure rates in blocking harmful prompts. Shocking? Not really. AI innovators prioritize utility, not security. Always have, always will. Security is an afterthought—bolted on later, rarely baked in from the start. And honestly, that’s fine. That’s how innovation works. If we waited for perfect security, we’d never move forward. But here’s the reality check: No AI system should be deployed without third-party security controls in place. Expecting models to self-regulate is effectively wishful thinking at best, negligence at worst. We’ve seen this story play out before. The internet, cloud computing, even mobile devices—every major tech leap started with a security Wild West before maturing (and even then, security still isn’t "solved"). AI is no different. So let’s not clutch our pearls when new models fail basic security tests. Let’s focus on what actually works: independent, external security layers that can adapt as fast as these models evolve. Innovation moves fast. Security needs to move faster. Follow us over at TrojAI for more hot takes. #CISO #CIO #Cybersecurity

🚨 AI Models Are Only as Safe as Their Weakest Prompt 🤖⚠️ AI is powerful, but is it behaving the way we expect? The real danger isn’t just AI getting things wrong—it’s AI being manipulated. 🔴 Jailbreaks and prompt injections can trick models into generating harmful content. 🔴 Subtle biases can go undetected and scale across millions of users. 🔴 Hackers can exploit AI models to behave in malicious ways. If we don’t red team AI models BEFORE deployment and monitor their behavior in real time, we’re flying blind. 🔍 How do we fix this? ✅ Pentest models like we would any critical system. ✅ Monitor AI inputs/outputs to detect manipulation attempts. ✅ Adapt in real time—AI security isn’t “set and forget.” AI isn’t magic—it’s just math. And bad math can be dangerous. Would you trust an AI model that wasn’t battle-tested? Let’s talk. ⬇️ #AIsecurity #RedTeaming #CISO #MachineLearning #CyberSecurity

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥  Reflecting on What Makes Us Stronger Last week, at our annual company kickoff, I felt a profound sense of pride as we celebrated our team's accomplishments. Since 2019, we’ve been on a mission to secure AI systems by focusing on protecting the integrity of model behavior. In an industry marked by significant growth and change, TrojAI addresses the unique security challenges that AI models introduce. New technology requires new and innovative solutions to secure it. This is why TrojAI is committed to pentesting and monitoring the models themselves to ensure they always behave as intended. In addition to our technology that focuses on securing model behavior, what sets us apart is the caliber of our team. They bring authenticity, depth, and relentless focus to every customer – no matter the scale. The passion they bring to work every day is truly inspiring, driving innovation and delivering exceptional results. Their commitment to understanding our customers' unique challenges ensures that we not only meet expectations but consistently exceed them. As we approach 2025, we do so as a team driven by purpose to build the best protections for AI models and applications. We are proud to earn trust as a favourite vendor one customer at a time. Follow us over at @TrojAISec for more hot takes. #Cybersecurity #GenAI #CISO #CIO #AI #HotTakeTuesdays #AISecurity

Cool!

"Leaders must embed cybersecurity at every stage of artificial intelligence (AI) adoption to safeguard sensitive data, ensure resilience and enable responsible innovation." 👈#CISO #CIO weforum.org/stories/2025/01/…

Incredibly excited to be invited to such a prestigious program with @Microsoft to secure AI models. #AISecurity #Cybersecurity #CISO #CIO prnewswire.com/news-releases…

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥 Breaking GenAI is Easier Than You Think @MsftSecIntel's latest whitepaper (lnkd.in/gHsJGdTC) highlights how generative AI systems can be vulnerable to simple attacks like prompt injections, jailbreaks, and others—echoing findings from important security research like Palo Alto Networks @Unit42_Intel's recent work on the Likert attack (lnkd.in/gSbkSQ8N). This research is critical for raising awareness about these threats, much like how naming vulnerabilities (e.g., Heartbleed or Shellshock) has helped shine a spotlight on them in the past. These attacks are novel and underscore the need for dedicated solutions to detect and prevent them. Security teams have an opportunity to focus on identifying and mitigating these vulnerabilities because these threats are already happening and evolving rapidly. While identifying vulnerabilities and publicizing new attacks play an essential role, the ultimate goal must be proactive detection and prevention. Malicious actors aren’t just working from what’s been published—and they aren't disclosing their techniques they use to exploit systems in the wild. The real challenge lies in staying ahead of what’s next by focusing on model behavior and creating robust, resilient defenses against both known 𝘢𝘯𝘥 unknown attacks. So, while publishing new jailbreaks is great for awareness, building resilient, proactive detection & prevention like what we have built at TrojAI is where we need to continue to focus. What’s your take? Are we focused enough on preventing these kinds of attacks now that we are seeing the increase in their viability? Let’s discuss! Follow us over at @TrojAISec for more hot takes. #Cybersecurity #GenAI #CISO #CIO #AI #HotTakeTuesdays #AISecurity

Want to help secure AI systems for the largest most innovative companies in the world? Our team is hiring! #Cybersecurity #AISecurity #DevOps linkedin.com/jobs/view/41310…

🔥 𝗛𝗼𝘁 𝗧𝗮𝗸𝗲 𝗧𝘂𝗲𝘀𝗱𝗮𝘆𝘀 🔥 OWASP Recognition: Proof That Following Your Instincts Pays Off When we started @TrojAISec, people said we were ahead of our time. Back then, there were no AI risk frameworks, no OWASP Top 10 For Large Language Model Applications & Generative AI, and barely any conversations about securing AI of any sort. Some thought we should wait for the market to catch up, while others claimed that the company with the most money would win. But we saw things differently. For us, it was never about timing or budgets—it was about solving critical problems before they became crises. We focused on what makes AI security uniquely AI security: protecting the integrity of AI behavior at build and runtime. We built automated pentesting and monitoring solutions trusted by Fortune 100 companies, staying true to our mission of safeguarding people and organizations in this new AI-driven world. Being recognized by @OWASP @LLM_Top10 is a deeply meaningful milestone. It reinforces what we’ve always believed: that innovation isn’t about waiting—it’s about acting on what you know to be right. Trust your instincts. Trust your purpose. If it’s rooted in protecting people and making a difference, it’s never too early to get started. Follow us over at @TrojAISec for more hot takes. #Cybersecurity #GenAI #CISO #CIO #AI #HotTakeTuesdays #AISecurity