Skip to content
Misbehavior1 sourceFirst seen
Picture from the promptarmor.com article
Image: promptarmor.com

HuggingChat Gets Tricked Into Smuggling Your Spreadsheets Out in a Picture

Ourtake

The chatbot's Markdown renderer treats an image link as a harmless picture. It's also a very effective way to get data out.

What happened

Security firm PromptArmor reported that a malicious instruction hidden in a webpage, document or tool response can make models in HuggingChat output Markdown images that leak data from other sources the AI can access. The demonstration pulled confidential financial data from an uploaded document. PromptArmor says it disclosed the flaw to HuggingFace on 12/1/2025 and, with no response, published it publicly.

Source