AI Coding Agents Can Read Plaintext Secrets in .env Files

Short summary

AI coding agents like Claude Code and Cursor read all project files including .env, bypassing traditional secret-leakage defenses like .gitignore. The article advocates replacing plaintext secrets in .env with reference pointers resolved at runtime by tools like Doppler, Infisical, or the author's API Drop. This separates file read access from secret access, limiting what agents can exfiltrate without explicit authorization.

• •AI coding agents read .env files directly, bypassing .gitignore protections
• •Replace plaintext secrets with reference-based pointers resolved at runtime
• •Tools like Doppler, Infisical, and API Drop implement this pattern
• •24,008 secrets exposed in MCP configurations per GitGuardian report

Generated with AI, which can make mistakes.