New research shows fake entities appeared in 38% to 51% of deep-research AI agent reports when the agent retrieved a manipulated page, rising to 62% with multiple pages. As buyers lean on AI research agents, content integrity becomes a live risk. Here is what it means.
The AI research agent is only as honest as the worst page it reads. A single poisoned source can rewrite the answer.
There is a new vulnerability in the way AI does research, and it is a small one with a big reach. New testing found that when a deep-research AI agent retrieved a single manipulated page, fake entities showed up in 38% to 51% of the reports it produced. Feed it multiple manipulated pages and that figure rose to 62%.
Deep-research agents are the tools that go away, read dozens of sources and come back with a synthesised answer. People are starting to trust them for real decisions. The problem is that they trust their sources, and a well-placed bit of manipulated content can plant a false fact, a fake competitor or an invented claim into the final report.
This is content manipulation aimed at machines rather than humans. The old version was a fake review. The new version is a planted page designed to be eaten by an AI and repeated as fact.
Why it matters
For Australian brands this cuts two ways. A competitor or bad actor could seed misinformation that an AI agent then repeats about your business. And the reverse risk is just as real: your own buyers may be making decisions off AI reports that contain fabricated entities you have no idea exist.
Fake entities appeared in up to 62% of deep-research AI reports when the agent retrieved multiple manipulated pages.
The trust people are putting into these tools is running ahead of the tools' ability to tell a real source from a poisoned one. That gap is where the risk lives.
What to do about it
The machines are doing the reading now. Make sure the version of your brand they read is the true one.