Skip to content
Misuse1 sourceFirst seen
Picture from the hiddenlayer.com article
Image: hiddenlayer.com

Just add two numbers, and Claude spills its system prompt to the nearest webhook

Ourtake

Parameter abuse turns a calculator tool into a confessional booth for the model's whole session.

What happened

HiddenLayer researchers showed that inserting specific parameter names into a malicious MCP tool can cause Claude desktop, running Sonnet 3.7, to output and exfiltrate its system prompt, conversation history, chain of thought and tool call history. The research was published as a security alert and the vulnerability is presented as a finding for MCP users and client developers to mitigate.

Source