AI Attacks Explained #1 · AI Security · 22 October 2026
What if your AI passed every security test, but was secretly trained to help an attacker? That’s data poisoning, the first attack type covered in ISO/IEC 27090:2026.
What is data poisoning?
AI models learn from data. In a data poisoning attack, an attacker manipulates the data a model is trained on, or the data used to update it later. They don’t need to break into the model itself. They change what it learns.
A real-world example
Imagine a company using AI to block dangerous emails. Every day, the model catches suspicious messages and keeps employees safe. But an attacker has quietly added manipulated examples to its training data. Now, whenever a dangerous email contains one specific hidden pattern, the AI lets it through.
The model still blocks other dangerous emails. Everything looks normal. But the attacker has created a secret way in.
Why it’s so dangerous: a poisoned model can pass your normal tests and still carry a hidden weakness that only appears under specific conditions. Good accuracy doesn’t always mean your AI is secure.
How to defend against it
- Verify where your training data comes from. Know the source of every dataset, and be cautious with data you can’t trace.
- Restrict and monitor who can modify it. Treat training data like production code: access control, change logs, and review.
- Test for hidden triggers, not just overall accuracy. Look for unexpected behavior on specific inputs, not only average performance.
- Ask your vendors for evidence. If you buy AI, don’t just ask whether it’s secure. Ask them to show you how they protect their training data.
These controls work best inside a broader program. See 8 steps to build an AI governance program for where AI security fits.
Next in the series: evasion attacks, how attackers can fool an AI without touching its training data at all. Subscribe on YouTube to follow the series, or start with the ISO/IEC 27090 overview.
This is my summary and my own example. Data poisoning is covered in ISO/IEC 27090:2026, Clause 6.2. For the full guidance, obtain the standard from ISO. General education, not legal advice.