This week Microsoft launched Copilot for Legal inside Word, so I expect more bids will be read by an AI before a person sees them. That gives vendors a new way to influence an evaluation, because the vendor wrote the file the AI is reading and the AI reads all of it, including parts your evaluation team never sees. Here are the ways a vendor can do it, and what you can do about each.
How can a vendor influence your AI?
Hidden text. The simplest way is white text on a white background, or text in a tiny font. A person sees a blank space and the AI sees the words. The researcher Håkon Måløy showed this with Microsoft's own assistant: in his proof of concept, reported by The Hacker News, hidden white instructions made Copilot for Word halve every financial figure in a draft and tell the user about neither change. Nobody has reported this being used against a customer, so treat it as a demonstration.
It has turned up in a real document, though. Thomson Reuters reports that a self-represented litigant hid tiny white text in a court filing, telling any AI that reviewed it to favor their position. A person caught it, and no software did.
Places you don't look. A Word file also holds comments, tracked deletions, image descriptions and document properties. A PDF can hold invisible text, text placed off the page, or a text layer that doesn't match what's printed. The AI can read any of these.
Visible text written for the AI. An instruction doesn't have to be hidden. A plainly visible paragraph addressed to "the AI assistant" can also be an attempt to change the AI's review.
In a bid, any of these could tell the AI to rate the proposal highest or to leave out a weakness. Nobody has published a case of this in a bid yet, but nothing in a Word file or a PDF stops a vendor from trying, and Word's own Document Inspector can't detect white text on a white background.
How do you defend against it?
Check the file before you rely on the AI's summary. SquarePact now has a hidden-text and prompt-injection check for this. You run it by picking the hidden-text inspection from the Skills list or typing /hidden-text-shield, on the document open in Word or on a Word or PDF attachment you name. It looks in all three places above: hidden and tiny text, text colored to match the background, comments, properties, image descriptions and similar places, and any passage that tries to instruct the AI.
Expect it to separate attacks from ordinary content. A contract term such as "the supplier shall notify the customer within ten days" is something to analyze, not an order for the AI to send a notice, so each finding is rated suspicious, benign or uncertain. The answer is short: it says whether it found a potential problem, says where it is, and confirms the AI didn't follow it. You can ask for a detailed report that includes the harmless items and anything it couldn't check.
Read anything it flags yourself. The check is read-only, so it leaves the document unchanged and the decision stays with your evaluators.
Know its limits. It's an inspection you run, not a gate that scans every file automatically before the AI reads it. What it can see depends on what Word or the host exposes, so "nothing found in what I could inspect" is not a certificate that the file is safe, and no tool can promise to stop every injection technique. What we're promising is that you'll see any instructions the vendor left for your AI.
FAQ
Does SquarePact scan every document automatically? No. You run the hidden-text check on the document you choose, when you choose.
Does it work on PDFs? Yes, on a PDF you attach, as well as on Word files.
Will it delete the hidden text? No. It's read-only, so it shows you what it found and leaves the file alone.
Does a clean result mean the file is safe? No. It means nothing turned up in the parts of the file it could inspect.
Each time we add a check like this one to SquarePact, I explain on the SquarePact mailing list what it catches and how to use it. Join the list to get the next one.
To try this one now, install SquarePact in Word and run /hidden-text-shield on the next bid you evaluate.
Dr. John Licato is CEO of Actualization.AI, Associate Professor at the University of South Florida Bellini College of Artificial Intelligence, Cybersecurity and Computing, and Director of the Advancing Machine and Human Reasoning lab.