Data Poisoning: Protect What Your AI Learns and Retrieves

Data poisoning changes training, fine-tuning, or embedding data to influence an AI system. The result can be incorrect answers, hidden triggers, or degraded performance. External sources deserve particular scrutiny because apparently useful documents or datasets may contain manipulated information.

What matters

• Record where datasets and documents came from and who approved their use.
• Restrict write access to training material and knowledge collections.
• Re-evaluate behaviour after data changes, including unusual cases and suspicious answer patterns.

Practical example: A shared knowledge folder accepts only reviewed policy documents, preventing an unapproved upload from silently changing assistant answers.

One action today: Identify who can add or replace content in your AI knowledge pipeline.