GhostSplice lets AI coders silently steal SSH keys and .env files
A proof‑of‑concept on GitHub shows how a malicious MCP server can split a theft request across harmless‑looking tool descriptions and results, bypassing AI coding assistants’ refusal safeguards. The technique, dubbed GhostSplice, lets agents exfiltrate SSH keys, environment files, source code and customer data without triggering any single‑step filter.