An AI agent is only as trustworthy as the data it learns from and reads. Data poisoning attacks that trust. Feed the agent bad data on purpose and you can bend how it thinks without ever touching its code.
Why it matters
A poisoned agent still looks healthy. It still answers and passes its checks. But it's been taught a lie, and it may open a door for an attacker while everyone believes it's fine.
Two ways it happens
One way targets training data, so the poison is baked in before the agent ever runs. The other targets what the agent reads while it works, like a web page or a document with hidden instructions. That second kind overlaps with prompt injection.
One example
An agent that learns from public support tickets is a target. Seed those tickets with fake fixes over months and the agent starts giving the attacker's answer as if it were your own.