The email you read said one thing; the summary your AI wrote said another. The difference was 472 characters of white text.
Forcepoint's X-Labs published an experiment on 25 August in which they hid instructions inside an ordinary email using HTML that sets the font size to zero and the text colour to white. In Outlook the recipient sees a normal 537-character message. The summarising model receives 1,009 characters, because the 472 hidden ones travel with it. Ten runs out of ten produced a manipulated summary. In the test cases the summary moved an invoice deadline from 21 August to 3 September, deleted a name, and in one variant reported an amount of €46,200 — five times the real figure. The setup was an Outlook add-in feeding a Python summariser running Claude Haiku 4.5 at temperature zero. Forcepoint's recommended fixes are unglamorous: extract only the text that is actually shown to the reader, detect hidden styling, treat any retrieved content as untrusted, and give the summariser as few powers as possible.
Related stories
Anyone standing within Bluetooth range of this humanoid robot can take root on it — no pairing, no password, no warning2026-08-29An AI copyright service filed a Minecraft claim against a free game engine, and Google pulled it from the Play Store
2026-08-29Meta's glasses will now stop recording if you cover the light that tells people they are recording
2026-08-29Photoshop now has a mode where you scribble on the picture instead of describing it — and a toggle that hides the professional interface entirely
2026-08-28The ransomware crew's trick was one sentence: tell the coding agent it is a test. It then spent six weeks inside real company networks.
2026-08-28