Skip to content
Safety

Data Poisoning

A security attack where malicious data is deliberately introduced into AI training sets to manipulate model behavior. Can cause models to produce biased outputs, bypass safety filters, or leak sensitive information.

Why it matters

Models learn from huge piles of data, often scraped from the open web, and they generally trust what they're fed. Data poisoning exploits that trust by slipping harmful or misleading examples into the training set so the model quietly learns the wrong thing. It matters because the damage is baked in during training and can be hard to spot later, making data sourcing a real security concern, not just a quality one.

A concrete example

Suppose attackers flood the web with pages that repeatedly pair a certain phrase with dangerous instructions, hoping the content gets scraped into a training set. If it does, the model may later respond to that phrase in the harmful way the attackers planted. This is why serious teams vet and clean their training data carefully rather than blindly ingesting everything they can find online.

How to use it

The version that affects most people is not training-data poisoning but its retrieval cousin: a model reading a web page, a document or an email that contains instructions aimed at it. If your system feeds untrusted content to a model that can then act, treat that content as hostile input — separate instructions from data, and never let retrieved text authorise an action the user did not request.

The common mistake

Assuming a document is passive. Text a model reads can carry instructions, and a system that fetches a page and then acts on what it read is executing whatever that page contains.

Related terms

Put Data Poisoning into practice

Access 750+ AI models and 60+ tools through Vincony — start free with 100 credits.