Executive brief
Upsonic, a framework for building AI agents, contains a vulnerability in how it handles Model Context Protocol (MCP) tasks. An attacker can use specific commands like 'npm' or 'npx' with malicious flags to bypass security filters and execute unauthorized commands on the server. This could allow a remote attacker to take full control of the system running the Upsonic process, potentially leading to data theft or service disruption.
Technical details
Upsonic 0.71.6 is vulnerable to command injection (CWE-77) within its MCP server/task creation functionality. While the application implements an allowlist for commands, it fails to properly sanitize arguments for permitted binaries such as 'npm' and 'npx'. These binaries support flags that allow for the execution of arbitrary OS commands. A remote attacker can define a crafted MCP task with malicious arguments to achieve remote code execution with the privileges of the Upsonic process. The vulnerability is addressed in version 0.72.0, which includes warnings regarding Stdio server command execution.
Affected products
- Upsonic upsonic < 0.72.0
Timeline
- 2026-04-15: disclosed
- 2026-04-15: advisory
- 2026-04-15: patched: Version 0.72.0 released