Data poisoning means slipping malicious content into the data an AI learns from or draws on to answer, in order to distort its behaviour: wrong answers, disinformation or a “backdoor” triggered by a specific word. It can target training: in a study published on 9 October 2025 by Anthropic, the UK AI Security Institute and the Alan Turing Institute, 250 poisoned documents were enough to plant a simple backdoor in models ranging from 600 million to 13 billion parameters, whatever their size. It can also target the document base of a RAG system: with PoisonedRAG, presented at the USENIX Security 2025 conference, researchers obtain the answer of their choice 90% of the time by adding five poisoned texts per targeted question to a base of several million texts. OWASP ranks it among the ten major risks of LLM applications, and the AI Act provides, where appropriate, for measures against data poisoning in high-risk AI systems (Article 15), applicable from December 2027 or August 2028 depending on the system.
In practice at Gensai
In its RAG projects, Gensai works from sources chosen and validated with the client, keeps track of where each document comes from and restricts write access to the document base.
Can a company document base be poisoned?
Yes, as soon as it includes content the company does not control: web pages, emails, documents sent by customers or partners, shared spaces that many people can edit. You need to know where each document comes from, who can add new ones, and be able to remove a dubious source quickly. In its threat overview published on 4 February 2026, ANSSI states that no case of training data poisoning has yet been brought to its attention, but considers that the risk exists.
Common mistake
Thinking that volume offers protection: “with thousands of documents, a few poisoned files won't change anything”. The studies quoted show the opposite: what matters is the number of malicious items, not their proportion, and a handful of well-placed texts can be enough to distort an answer.
Same theme
Definition written with AI assistance, checked by Sébastien Lecocq, co-founder of Gensai.
Going beyond the definition?
From concept to project: Gensai builds custom, sovereign, GDPR-compliant AI solutions.
Related services: Chatbots and answer engines, Custom AI solutions