I’ve seen a lot in Silicon Valley. I’ve built and sold two companies, RemoteTeam to Gusto and MovieLaLa to Gfycat. I’ve been lucky enough to be an early investor in over 200 companies, including some of the names you see in the headlines every day like Anthropic, OpenAI, Scale AI, and Hugging Face. But after analyzing over a hundred AI threat incidents in the last year, I’ve seen a pattern that genuinely scares me. And I’m not easily scared.
I recently sat down with 50 hackers—the kind of people who can find a way into almost any system. I wanted to understand how they think about AI threats. Their answers were terrifying. They’re not just exploiting code anymore; they’re exploiting the very nature of intelligence. And most companies are completely unprepared for what’s coming.
This isn’t another high-level, abstract post about the “dangers of AI.” This is a tactical guide for founders, engineers, and anyone who builds or protects digital products. This is what you need to know before it’s too late.
The New Attack Surface: Your AI Itself
For years, cybersecurity has been about protecting networks, servers, and databases. We built firewalls, we encrypted data, and we trained employees not to click on suspicious links. That’s all still important, but it’s not enough. The new front line in cybersecurity is the AI model itself.
Think about it. Your AI models are making decisions. They’re classifying data, predicting outcomes, and even generating content. What happens when an attacker can manipulate those decisions? What happens when they can poison your training data or trick your model into revealing sensitive information? It’s not a theoretical risk. It’s happening right now.
I saw one case where a startup was using an AI model to detect fraudulent transactions. A sophisticated attacker was able to slowly “teach” the model that certain types of fraudulent transactions were legitimate. They did this by feeding it a steady stream of carefully crafted data. By the time the startup realized what was happening, they had lost millions.
The Three Main Types of AI Threats
When I talked to hackers, they all mentioned three main types of attacks. These are the things that keep them up at night—and they should keep you up at night too.
1. Data Poisoning
This is the one that scares me the most. Data poisoning is when an attacker intentionally pollutes your training data. This can be incredibly difficult to detect, and the consequences can be catastrophic. Imagine an AI model used for medical diagnoses that has been poisoned to misdiagnose certain types of cancer. The potential for harm is enormous.
One of the hackers I spoke to told me about a time they were hired to test the security of a self-driving car company. They were able to poison the training data for the car’s computer vision system. They did this by subtly altering a small percentage of the training images. The result? The car started misidentifying stop signs as speed limit signs. It was a terrifying demonstration of how easy it can be to cause chaos.
2. Model Evasion
Model evasion is when an attacker finds a way to trick your AI model into making the wrong decision. This is often done by creating “adversarial examples”—inputs that are specifically designed to fool the model. For example, an attacker could create an image that looks like a cat to a human but is classified as a dog by an AI model.
This might sound like a fun academic exercise, but the real-world implications are serious. I’ve seen cases where attackers have used adversarial examples to bypass spam filters, fool content moderation systems, and even trick facial recognition systems. One of the hackers I interviewed showed me how he could print a special pattern on a t-shirt that made him invisible to a popular security camera system. He could walk right past the camera, and it wouldn’t even register that he was there.
3. Model Inversion and Extraction
Model inversion and extraction are when an attacker is able to steal your AI model or the data it was trained on. This is a huge risk, especially for companies that have invested heavily in developing proprietary models. If an attacker can steal your model, they can use it for their own purposes, or they can sell it to your competitors.
I talked to one founder who had spent years and millions of dollars developing a cutting-edge AI model for drug discovery. An attacker was able to use a model extraction attack to steal the model. The founder didn’t even know it had happened until he saw a knock-off version of his product on the market a few months later. It was a devastating blow to his business.
How to Protect Yourself
So, what can you do to protect yourself from these new types of threats? It’s not easy, but it’s not impossible. Here are a few things I recommend.
- Secure your data pipeline. The first step is to make sure that your training data is secure. This means implementing strict access controls, monitoring for suspicious activity, and using data provenance techniques to track the origin of your data.
- Use robust training methods. There are a number of training methods that can help make your models more robust to attack. These include adversarial training, which involves training your model on adversarial examples, and differential privacy, which adds noise to your data to make it more difficult for an attacker to extract sensitive information.
- Monitor your models in production. Once your models are in production, you need to monitor them for signs of attack. This means tracking their performance, looking for anomalies, and using tools to detect adversarial examples. I’ve invested in several companies that are building tools specifically for this purpose.
- Have a response plan. No matter how well you protect yourself, there’s always a chance that you’ll be attacked. That’s why it’s so important to have a response plan in place. This plan should outline who is responsible for what, how you will communicate with your customers, and what steps you will take to mitigate the damage.
The Future is Here, and It’s Dangerous
I’m not trying to be an alarmist. I’m a builder and an optimist at heart. I believe that AI has the potential to solve some of the world’s most pressing problems. But we can’t be naive about the risks. The same technology that can be used for good can also be used for evil.
We’re at a critical juncture in the history of technology. The decisions we make today will have a profound impact on the future. We need to build AI systems that are not only powerful but also safe, secure, and aligned with human values. It’s a tall order, but I’m confident that we can do it. We have to.
Frequently Asked Questions
How often is this guide updated?
I revisit and update my guides regularly as I learn new things and as the market evolves. The core principles tend to stay stable, but specific tactics and tools get refreshed based on what's working right now.
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.
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.
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.