The attacks that target AI systems and the controls that reduce them.
OWASP's Top 10 is a widely used checklist of risks in applications built on large language models. Here is what each entry means and what changed between the 2025 and 2026 editions.
Prompt injection tops every list of LLM risks. Learn the difference between direct and indirect injection, why national agencies say it may never be fully fixed, and what limits the damage.
Some attacks on AI target the model itself rather than the prompt. This article maps data poisoning, training data extraction and model theft to the lifecycle stage each one exploits.
Adversarial examples fool a model's judgement; jailbreaks talk a model out of its safety rules. Both exploit the same weakness, and neither has a complete fix. Here is how defenders should think about them.
Downloading a model can run someone else's code on your machine. This article explains where AI supply chain risk sits, why pickle-based model files are dangerous, and which controls actually help.
If prompt injection cannot be fully prevented, the system around the model has to contain it. This article covers excessive agency, the lethal trifecta and the design patterns that limit damage.
Several frameworks describe AI attacks and defences, each for a different job. This article explains what NIST AI 100-2, MITRE ATLAS, OWASP, Google SAIF and the joint agency guidelines are for, and how they line up.