Data poisoning
An attack that corrupts the data a model learns from or relies on, to change its behaviour.
Data poisoning is an attack in which an adversary inserts or modifies data that a model is trained or fine-tuned on, or that it retrieves at run time, to change its behaviour, for example to plant a hidden trigger or degrade accuracy. NIST's taxonomy of adversarial machine learning classes poisoning as an attack on the training stage. For financial agents, poisoning risk extends to the knowledge bases and supplier data they consult, so data sources need provenance and change control.
Agent Minute explains this term on 27 December 2026.
Related terms
Indirect prompt injectionPrompt injection hidden in content an AI system retrieves, such as a web page, document or email.Red-teamingStructured adversarial testing of an AI system to find flaws, harmful behaviours and vulnerabilities before attackers do.Model driftA decline in a model's or agent's performance over time as data, users or conditions change.Third-party risk managementManaging the risks of relying on outside providers across the relationship, from due diligence to exit.