noun
AI model poisoning is a type of cyberattack where attackers deliberately corrupt an AI system by feeding it misleading or malicious data. This can happen during training (data poisoning) or after deployment (adversarial attacks). The goal is to make the AI produce incorrect, biased, or harmful outputs.
The attack works by exploiting the AI’s reliance on large datasets. For example, if an image-generating AI is trained on altered photos of cats labeled as ‘dogs,’ it may later misclassify real cats. Similarly, a chatbot fed false historical facts could spread misinformation.
For developers, this means extra security checks are needed. Companies typically use data validation, anomaly detection, and adversarial testing to protect their models. A real-world case saw a self-driving car AI misread stop signs altered by tiny stickers, showing how subtle changes can cause big failures.
A hacker sneaks fake product reviews into a recommendation AI, making it promote low-quality items.