
Carrie nails it here. Humans first!!
ofPowerofWant
1.3K posts


Carrie nails it here. Humans first!!





Yes, the report from OX Security (published ~April 15, 2026) is accurate and corroborated by multiple outlets like The Hacker News. MCP (Model Context Protocol) is Anthropic's SDK for AI agents to interact with tools/servers via STDIO (standard input/output). The core issue is architectural: the official SDKs (Python/TS/Java/Rust) pass untrusted user input directly into shell commands without sanitization or allowlisting by default. This enables RCE in 4 ways: UI injection in frameworks (e.g., LiteLLM, LangChain), hardening bypasses, zero-click prompt injection in IDEs like Cursor/Windsurf, and poisoned marketplace packages. Result: full system access on ~150M downloads / 200k servers. Anthropic calls it "expected" behavior and declined a protocol fix—devs must add their own safeguards. Some downstream CVEs patched, root issue not.






