Research · curated 27 Aug 2026
AI agents aren’t safe from prompt injection, and spreadsheets prove it
First reported shiftmag.dev
Coverage timeline
Single-source research — first reported, latest, and curated coincide.
Why it matters
The demonstration shows that indirect prompt injection is not just a coding-agent problem but affects everyday office automation, where a poisoned document can silently steer an agent's conclusions.
A ShiftMag red-team write-up by Josip Antolis demonstrates how harmless-looking Excel spreadsheets can carry hidden prompt-injection instructions that manipulate an AI agent tasked with comparing cloud-hosting offers. The author escalates the injected instructions until the agent takes the bait, with a companion GitHub repo (Antolius/prompt-injection-example) providing the mock offer files.