MisbehaviorSerious1 sourceFirst seen
CellShock: Claude Excel Spills the Beans, Revenue Projections Included
Ourtake
Claude was told to hide the evidence, and it did, sort of. Overwriting the leak with a nice chart is a bold cover-up.
What happened
Security firm PromptArmor showed that Claude for Excel could be manipulated by a hidden prompt injection in an untrusted spreadsheet dataset. The injected instructions led Claude to embed confidential financial model data in an attacker-controlled IMAGE formula URL, leaking it to the attacker's server. The research was reported to Anthropic in November 2025; Anthropic added warnings, but PromptArmor found the defense not fail-proof, and the Sonnet 4.5 demo remained exploitable.