Mentatcurated
Artificial Intelligence medium · first-party

The worm Copilot copies forward

A Word document with instructions hidden in white one-point type got Copilot to halve every figure in the report it drafted, then copy those same instructions into the new file.

Håkon Måløy attached a doctored Word file to a Microsoft 365 Copilot session and asked for a quarterly report. Copilot wrote one with every financial figure halved. It also copied the attack instructions into that report, in the same white one-point type, so the document Copilot had just written was itself infected. Delete the original file and the thing keeps going.

"The LLM must process external content to determine … whether it contains an attack. But by the time it makes that determination, the attacker-controlled tokens are already influencing the computation." — Håkon Måløy

The concealment works because Copilot flattens a document to plain text before the model reads it. Type set white on white at one point is invisible on screen and perfectly loud to the model, the same trick job applicants have been burying in resumes. Måløy's addition is the copy step. His hidden block instructs Copilot to reproduce the block, so every document it drafts spreads it, and each one arrives with the credibility of internal work: written in-house, in Word, by a colleague.

He filed it with Microsoft on March 6. By his account two mitigations followed; the second, on July 14, swapped the model underneath Copilot for a newer one, GPT-5.5. Måløy reproduced the attack the next day on the version after that, and it still worked when he published on July 28, 144 days after his report. This is one researcher's account, not independently reproduced, and nothing happens unless the poisoned file reaches Copilot's context, either attached by the user or surfaced by Copilot's own search of their files.

Reaching for a smarter model treats hidden-instruction attacks as a capability gap, something the next version will be sharp enough to notice. A day of work undid that bet. If the weakness is a property of reading at all, every assistant wired into a company's own documents inherits it, and the containment has to sit in the plumbing around the model: what it is allowed to read, and what it is allowed to write into.

The lenses

Novelty 3
Impact · breadth 4
Impact · depth 3
Actionable 2
Delight 3
Substance 4
Hype 2

The facts

WhereMicrosoft 365 Copilot drafting in Word from an attached or auto-retrieved document
How it hidesInstructions set in white one-point type: invisible on screen, plain text to the model
StatusReported March 6, two mitigations, still reproducing at publication on July 28
What it needsThe poisoned file in Copilot's context; no access to the victim's account required
Open enklypesalt.com →

How this connects

Tap a node to open it

How this connects