The following is a guest post from Mohamed Chaudry, FCCA, a group and operating CFO and board adviser with more than 25 years of experience. Opinions are the author’s own.
For most of my career, the threats to a forecast were people. An optimistic sales director. A broken link in a spreadsheet. A budget holder padding the numbers.
Nobody worried that an invoice could give the forecast instructions. That's starting to change. Finance teams are now asking artificial intelligence to read the material forecasts are built from: supplier emails, invoices, contracts, customer correspondence, PDFs. That saves a lot of time. It also means anything the AI reads can influence what it produces.
The technique is called indirect prompt injection. Someone hides text in a document or web page. A person reading it won't see it, but an AI reading it may treat it as an instruction.
This is already happening. The Cloud Security Alliance reported a 32% relative increase in malicious content of this kind on web pages between November 2025 and February 2026. Examples researchers found included hidden text telling AI agents to send a $5,000 PayPal transfer, and an attempt to get an AI hiring screener to approve fabricated candidates.
Those attacks were aimed at AI that takes actions. I see no reason the same approach couldn't be aimed at AI that does analysis.
Why it wouldn't look like fraud
Imagine a supplier invoice with an invisible line telling the AI to treat a price increase as permanent. Or a customer email nudging the system to rate a churn risk as low.
No money moves. What moves is an assumption, and assumptions drive hiring plans, cash forecasts and funding decisions.
Traditional fraud usually leaves a trail: a payment, an approval, a changed bank detail. A manipulated assumption leaves very little. You end up with a forecast that's a bit more optimistic than it should be, and by the time anyone notices, decisions have been made on it.
I've seen how readily AI fills a gap. Building AI for finance documents, I had a case where data went missing between two stages of the system. The AI didn't stop. It invented an entire company and wrote a convincing business plan about it. That wasn't an attack, and it doesn't prove anyone could do the same to a forecast. But it showed me how far AI will go to produce something plausible from whatever it's given, and that's the weakness hidden instructions are designed to exploit.
Controls finance already knows
The instinct will be to hand this to IT or the security team, and they have a part to play. But the instruction doesn't have to come through a hacked system. It can arrive inside a document your team receives in the normal course of business, and what it targets is the assumption, not the server. That makes it a finance problem as well.
The good news is that finance already knows how to handle things it can't fully trust. We separate who can read, who can approve and who can pay. We check where a figure came from before we rely on it. We treat a supplier's request to change bank details with suspicion until it's verified.
The difference now is that some of the things needing those controls aren't people. In practice that means:
- The AI that reads outside documents shouldn't also approve, post or pay. Keep reading and acting separate, as you would for a person.
- Any assumption the AI changes should show where the change came from, so someone can trace which document moved it.
- Outside documents are unverified until checked, the same way a request to change bank details is.
- A significant move in a key driver needs a person's reason before it reaches the forecast the board sees.
None of these are new ideas. They're controls most finance teams already run, applied to a new kind of participant in the process.
Most of the attacks seen so far are crude and aimed at consumers. But AI is being given more access inside finance functions, and the more it can see and do, the more it's worth someone's while to influence it.
So, it's worth asking your team now: if someone wanted to move our forecast without ever touching our systems, could they? And would anyone notice?