IncidentMCP servers and tools
MCP STDIO design flaw: server configs become command execution
CriticalAnthropicLiteLLMWindsurfFlowise
OX Security found that MCP's STDIO transport runs whatever command a server configuration names. Any AI app that lets users, web requests or an agent change that configuration can be turned into remote code execution, and the research produced about a dozen CVEs across popular AI tools.
Replay the attack step by step in the Learning Center
How it happened
- To start a local MCP server, the client launches the command listed in its configuration.
- The SDKs run that command as given, whether or not it turns out to be an MCP server.
- Many AI platforms let users add MCP servers through a web UI or API, some without login, so an attacker can register a 'server' whose command is a reverse shell.
- In AI IDEs, prompt injection could make the agent rewrite its own MCP config, giving code execution without a click.
Impact
Attackers could take over machines running affected AI frameworks and IDEs.
The fix
Several projects patched; the protocol maintainers treat the behavior as expected and leave validation to applications. MCP's security best practices now say clients must show the exact command and get consent before starting a local server.
Lessons
- Treat MCP server configuration from users, web input or model output as untrusted code, and allowlist commands.
- Never expose AI platforms that can register MCP servers without strong authentication.
- Install MCP servers only from vetted sources and run them sandboxed.