Forget Everything You Know About AI Security. The Rules Have Changed.
A few weeks ago, a founder I mentor called me in a full-blown panic. His fintech startup, a company I genuinely believe is on a rocket ship trajectory, was targeted by a hyper-realistic AI phishing attack. This wasn't the clumsy, typo-ridden email from a supposed Nigerian prince that we all learned to spot a mile away. This was a masterclass in deception.
The email appeared to be from his CFO, referencing a confidential project by its internal codename. The tone was perfect, capturing the CFO's usual directness and sense of urgency. It even mentioned a specific detail from a meeting they'd had the day before. The only reason the scam unraveled was pure, dumb luck. The employee who received the email was about to wire a significant sum of money when the real CFO happened to walk past her desk and asked what she was working on. Crisis averted, but it was a terrifyingly close call.
This is the new battleground. If your company's security strategy is still centered on building bigger walls and teaching employees to spot obvious fakes, you're not just fighting a war that ended a decade ago. The explosion of artificial intelligence hasn't just created new markets and opportunities; it has unleashed a new and terrifying class of vulnerabilities. And from what I’ve seen across my portfolio and beyond, most companies are dangerously unprepared.
I’ve been in the Silicon Valley trenches for over twenty years, first as a founder with two exits under my belt, and now as an angel investor in over 200 companies. I’ve had a front-row seat to the rise of cloud, mobile, and SaaS. But nothing—absolutely nothing—has moved with the speed and disruptive force of AI. My investments in foundational companies like Anthropic, OpenAI, Scale AI, and Hugging Face have given me a unique perspective on both the incredible promise and the hidden perils of this technology. I’m not a doomsayer, but I am a realist. And the reality is, if you’re ignoring AI security, you’re not just taking a risk; you’re being negligent.
The Old Security Playbook Is Officially Obsolete
For the better part of two decades, we operated under a security paradigm that was, in hindsight, quite simple. We built digital fortresses with firewalls, we ran drills to teach our teams to recognize suspicious links, and we deployed armies of antivirus software to hunt for known threats. It was a game of cat and mouse, but the rules were generally understood. AI has thrown that rulebook into a bonfire.
Adversarial AI is one of the most chilling examples. This isn't about brute-force attacks; it's about subtle, surgical manipulation. An attacker can introduce almost imperceptible "noise" into an image, a video, or a block of text that is invisible to the human eye but completely fools a machine learning model. Imagine a self-driving car’s perception system being tricked into seeing a 35 mph speed limit sign as an 85 mph sign because of a few strategically altered pixels on a sticker. Or a medical imaging AI that misdiagnoses a malignant tumor as benign because of a tiny, adversarial perturbation in the scan data. These aren’t science fiction scenarios. Researchers have been demonstrating these vulnerabilities for years, and now, these techniques are moving from the lab to the wild.
Then there’s the weaponization of AI for social engineering, like the AI phishing attack I described. The ability of large language models to generate fluent, context-aware, and highly personalized text is a goldmine for attackers. They can scrape LinkedIn for your professional background, your company’s blog for project details, and social media for your personal interests. Then, they can craft an email that doesn’t just look like it’s from your boss; it sounds like it’s from your boss, referencing the project you just discussed and asking for the file you’re working on. The same technology can be used to clone a voice with just a few seconds of audio, allowing an attacker to leave a believable voicemail or even engage in a real-time conversation. The era of easily detectable phishing is over.
My New Playbook for the AI Era
So, how do we fight back in a world where we can no longer trust our own eyes and ears? We have to fundamentally shift our mindset from defense to resilience. We must operate under the assumption that attackers are already inside our systems and design our defenses accordingly. This is the playbook I drill into my portfolio companies.
Verify Everything, Relentlessly. This is the core tenet of the Zero Trust security model, and it’s no longer optional. The old model of a trusted internal network and an untrusted external world is dead. In the age of AI, every user, every device, every application, and every API call must be treated as a potential threat. Identity must be continuously verified and authenticated, regardless of where the request originates. There is no "inside the firewall" anymore.
Embrace Radical Least Privilege. The principle of least privilege—giving a user or system only the access it absolutely needs to perform its function—is not new. But it needs to be applied with a new level of rigor. An AI model designed for sentiment analysis should not have access to personally identifiable information. A marketing automation agent should not have the ability to modify production code. By radically restricting the "blast radius" of every component in your system, you limit the damage an attacker can do if one of those components is compromised.
Assume You Are Breached. This is the most important mental shift. Stop thinking about security as a way to prevent breaches and start thinking about it as a way to survive them. This means having robust systems for detecting anomalous behavior in real-time. It means having a well-rehearsed incident response plan. And it means having immutable backups and the ability to quickly restore your systems to a known-good state. When the breach happens—and it will happen—your ability to detect it, contain it, and recover from it will be what separates a minor incident from an extinction-level event.
The Tools I’m Betting On
Principles are the foundation, but you can’t build a house without tools. The good news is that a new wave of security companies is rising to the challenge, building AI-powered defenses to combat AI-powered threats. Here are a few of the platforms I’m consistently recommending to my founders:
Cycode: I’m incredibly bullish on what Cycode is doing. They’ve built a unified platform that brings together all the critical pieces of application security—from static analysis (SAST) and software composition analysis (SCA) to infrastructure-as-code (IaC) scanning and container security. Their "Context Intelligence Graph" is the secret sauce, mapping the entire software development lifecycle to understand how a vulnerability in a single library could be exploited in production. Their AI Exploitability Agent is a force multiplier for security teams, automatically determining which of the thousands of alerts are actually exploitable, allowing developers to focus on what matters.
Snyk: Snyk has done a phenomenal job of building a security platform that developers actually like to use. Their focus on the developer experience is key to their success. They provide tools that integrate seamlessly into the developer’s workflow, finding and fixing vulnerabilities in code, open-source dependencies, and containers. Their DeepCode AI, which combines symbolic AI with generative AI, is incredibly effective at finding complex vulnerabilities and providing actionable remediation advice.
Checkmarx One: For larger organizations with a sprawling and complex application landscape, Checkmarx One is a powerful solution. It’s a comprehensive platform that provides a single pane of glass for SAST, DAST, API security, and more. I’m particularly impressed by their investment in agentic AI assistants, which can autonomously hunt for and even help remediate threats across the entire development lifecycle. This is the future of security operations.
A Story From the Trenches
Let me give you a concrete example of this new playbook in action. One of my portfolio companies, a startup in the personalized medicine space, was the target of a sophisticated data poisoning attack. The attackers were attempting to subtly corrupt the training data for their diagnostic AI, hoping to skew its results and undermine the credibility of their platform. An attack like this would have been almost impossible to detect with traditional security tools.
But this company had built its security strategy around the principles I’ve outlined. Their implementation of Zero Trust meant that the attacker’s initial foothold was contained. Their use of least privilege access prevented the attacker from moving laterally across their network. And their AI-powered security tools were the real heroes. Cycode’s platform detected the anomalous data being fed to the model, Snyk identified the vulnerable open-source library that the attacker had exploited to gain access, and Checkmarx helped their team quickly patch the vulnerability and validate the fix. The attack was neutralized before it could have any impact on their diagnostic results.
This is the power of a modern, AI-centric security strategy. It’s not about building an impenetrable fortress. It’s about building a resilient, adaptable system that can withstand and recover from the inevitable attacks.
The Road Ahead
The AI revolution is just getting started, and the security challenges are only going to become more complex. We’re going to see the rise of fully autonomous AI agents, both for good and for ill. We’re going to see attackers using AI to discover and exploit zero-day vulnerabilities at a speed and scale that we can barely imagine. And we’re going to have to grapple with the profound ethical and societal implications of this technology.
But I’m an optimist. I believe that for every new threat that AI creates, it also provides us with new and more powerful tools to defend ourselves. The key is to be proactive, to be vigilant, and to be willing to throw out the old playbook and embrace a new way of thinking about security.
So, what should you do tomorrow? Start by having an honest conversation with your team. How would you have fared against the phishing attack I described at the beginning of this article? Do you have a clear picture of all the AI tools and models being used in your organization? Have you implemented the principles of Zero Trust and least privilege access? If the answer to any of these questions is no, then you have work to do. The future of your company, and the trust of your customers, depends on it.
Frequently Asked Questions
How should I work through this guide?
Don't try to absorb everything in one sitting. Read through once to get the big picture, then go back and work through each section as it becomes relevant to your current challenges. Bookmark it and return to it regularly.
Who is this guide designed for?
This guide is written for founders and operators who want practical, actionable advice rather than theoretical frameworks. Whether you're just starting out or scaling an existing business, the principles here apply across stages.
What if I disagree with some of the advice?
Good. That means you're thinking critically, which is exactly what a good founder should do. Take what resonates, test it, and discard what doesn't work for your specific situation. No advice is universal.