I lost sleep over Adversarial AI for months. It wasn’t the kind of restless night you get before a big launch. It was a deep, gnawing anxiety, the kind that comes from seeing a crack in the foundation of everything you’re building.
My name is Sahin Boydas. I’ve been in the Silicon Valley trenches for a while now. I’ve built and sold two companies, RemoteTeam to Gusto and MovieLaLa to Gfycat. I’ve been fortunate enough to write angel checks for over 200 companies, including some names you might recognize like Anthropic, OpenAI, and Scale AI. I’ve seen a lot. But what I’m seeing now with AI is different. There’s a story being told about the dangers of AI, and it’s the wrong one. I’m here to tell you the truth.
Forget the pictures of turtles that an AI thinks is a rifle. That’s a parlor trick. It’s the kind of thing academics write papers about, but it doesn’t capture the real, business-ending threat that’s brewing. The real danger isn’t about fooling a model; it’s about manipulating reality, and it’s already happening.
The Wake-Up Call I Can't Ignore
About a year ago, I was on a board call for one of my portfolio companies—a fast-growing e-commerce startup. Their core advantage was a dynamic pricing engine that adjusted prices in real-time based on dozens of signals. It was a thing of beauty. And it was slowly killing them.
Their margins were eroding. Week after week, the numbers were just a little bit off. Not enough to set off major alarms, but enough that the CFO was starting to look pale. We dug into it. We blamed the market, seasonality, a new competitor. We blamed everything except the one thing we trusted implicitly: the algorithm.
It turned out a competitor hadn’t hacked our systems. They had poisoned our data. For months, they had been feeding our model subtly manipulated signals through a network of seemingly legitimate user accounts. They made certain products look less desirable, others more so. They didn’t break the model; they just nudged it, day after day, in a direction that favored them. By the time we figured it out, we had lost over $3 million in potential revenue. The model was working perfectly, but it was working on a lie.
This is the new reality of adversarial AI. It’s not a zero-day exploit. It’s a thousand-day campaign of subtle manipulation.
A New Playbook for an Adversarial World
The old security playbook is useless here. You can’t just patch a vulnerability or update your firewall. You have to fundamentally change how you think about building and deploying AI. This is the playbook we developed after that incident, the one we’re implementing across my portfolio.
1. Embrace Paranoid Data Provenance
Your model is only as good as the data it’s trained on. You need to be absolutely paranoid about where your data comes from. Every single data point needs a clear lineage. Who created it? How was it collected? Who has touched it? We started treating our data pipelines with the same security rigor as our production code. If you can’t verify the source of your data, you can’t trust your model. Period.
2. Red Teaming on Steroids
Your security team is probably great at finding vulnerabilities in code. But can they think like a behavioral psychologist? A disgruntled employee? A nation-state actor? We started hiring outside-the-box thinkers—novelists, economists, even a former professional poker player—to red team our AI systems. Their job isn’t to find bugs in the code, but to find exploits in the business logic. They ask questions like, “How could I make this company fire its best salesperson?” or “How could I make this company unknowingly violate international sanctions?” It’s a different level of thinking.
3. Monitor the Business, Not Just the Model
Stop obsessing over model accuracy. Start obsessing over business outcomes. Your model’s accuracy could be 99.9%, but if that 0.1% is costing you millions, you have a problem. We now build real-time dashboards that track the second and third-order effects of our models. We look for strange correlations and unexpected changes in user behavior. We’re not just looking for a single compromised prediction; we’re looking for a compromised reality.
4. Build for Resilience with Zero-Trust AI
Assume your model will be compromised. It’s not a matter of if, but when. The question is, what happens then? A Zero-Trust AI architecture assumes any component, any model, any data source could be malicious. This means building in circuit breakers and fail-safes. If a model starts behaving erratically, can you automatically revert to a simpler, more stable version? Can you isolate the blast radius of a compromised system? Resilience, not just prevention, is the goal.
The Security Tools We Desperately Need
Here’s the hard truth: the tools for this new era of AI security barely exist. The market is flooded with products that promise to detect bias or explain model predictions, but very few are built to handle active, intelligent adversaries who are targeting your business logic. There’s a massive opportunity here for founders and investors.
We need a new generation of AI threat detection platforms that can identify these slow, subtle attacks. We need tools that can automatically verify data provenance at scale. We need to get serious about defending against zero-day AI exploits that target the very assumptions our models are built on. This is the next frontier of AI cybersecurity.
The Real Game Has Just Begun
For too long, we in Silicon Valley have been captivated by the potential of AI. We’ve focused on building bigger, faster, and more complex models. We’ve been playing a game of creation. But we’ve forgotten that with creation comes destruction.
The next unicorn won’t be a company that builds a slightly better AI model. It will be the company that figures out how to protect AI from the adversarial world it lives in. The game has changed. It’s time we changed with it.
Frequently Asked Questions
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.