From the vault
Credential leaks, MCP security, zero-knowledge architecture — written for engineers building with AI coding agents.
Many 'secrets for AI agents' MCP servers return the raw secret value straight into the model's context. Here's how to tell which ones do, why it matters, and the architecture that makes it structurally impossible.
Read post →Phase is open-source and end-to-end encrypted — but the moment you turn on the public API, an integration, rotation, or a machine identity, the server can decrypt. Here's what 'zero-knowledge' actually requires, checked against Phase's own public source.
Read post →A practical guide to preventing API keys from leaking into Claude Code, Cursor, and Cline — covering common failure modes, vault patterns, and incident response.
Read post →macOS Keychain keeps secrets encrypted at rest. The moment you use the security CLI to pull them into your shell — and Claude Code is running — you've already lost.
Read post →HashiCorp Vault, Kubernetes secrets, Doppler, Infisical, systemd — none of them were designed for a world where the process requesting credentials is an AI agent that can read its own environment.
Read post →Anthropic accidentally shipped 60MB of Claude Code's unobfuscated source code in an npm package. Here's what security researchers found inside — and what it means for how you think about AI agent credentials.
Read post →Cloudflare Workers secrets are designed for production deployment, not local development. Here's what they protect, what they don't, and how to close the gap when using Claude Code.
Read post →The advice spreading across AI communities — 'just create a .env file and have Claude read it' — is wrong, and here's why it matters.
Read post →Claude Code can read your project files, including .env. Here's the right way to handle API keys so they never enter the model's context window.
Read post →The standard .env approach exposes your API keys to Claude's context window. Here's the correct way to store and use secrets with Claude Code.
Read post →OpaqueVault is the only MCP-native secret manager built for AI coding agents. Here's how it works, why the architecture matters, and how to set it up in five minutes.
Read post →How to manage credentials across MCP servers without exposing them to the AI. Architecture, patterns, and a reference implementation with OpaqueVault.
Read post →A zero-knowledge secret manager means the server cannot decrypt your secrets. Here's why that architecture matters — especially when AI coding agents are in the loop.
Read post →Researchers found 24,008 live API keys and credentials inside Claude Desktop MCP configuration files. Here's what went wrong and how zero-knowledge vaults change the equation.
Read post →HashiCorp confirmed HCP Vault Secrets reaches end-of-life on July 1, 2026. If you're using it today, here's what you need to know and where to move your secrets.
Read post →Claude Code, Cursor, and GitHub Copilot dramatically accelerate development — and, if misconfigured, dramatically accelerate credential exposure. Here's the structural problem and how to fix it.
Read post →