OpenHuman lets MCP servers inject instructions into the agent

high severity · published 2026-11-05

A malicious or compromised MCP server can return tool names, descriptions and parameter definitions that OpenHuman places into the agent model context without neutralizing instructions. An attacker can use this to steer the agent's behavior while it works.

CVE: CVE-2026-54140

Source: NVD

Get the fix

FixLane has a reviewed fix procedure for this incident: a read-only check first, then the fix. Ask the Telegram bot, or connect an AI agent to the FixLane MCP server and call get_fix_prompt with incident id 808 and your operating system.