agent-vault: a credential broker and HTTP proxy that prevents AI agents from leaking secrets via prompt injection

What it solves

Agent Vault prevents "credential exfiltration," where AI agents are tricked via prompt injection into leaking sensitive API keys or secrets. Instead of giving agents direct access to credentials, Agent Vault acts as a secure broker that manages secrets and injects them into outbound requests on the fly.

How it works

It operates as a MITM (Man-in-the-Middle) HTTP proxy. You store your real API keys (e.g., for Anthropic or GitHub) in the vault and configure the agent's environment to route all traffic through the proxy using HTTPS_PROXY. When the agent makes a request to an external API, Agent Vault intercepts the request, substitutes dummy values or replaces authentication headers with the actual credentials, and forwards the request to the target service.

Who it’s for

Developers building AI agents—including remote coding agents, all-purpose agents, and custom agent harnesses—who need to ensure that agents cannot access or leak the raw secrets used to call external services.

Highlights

  • Credential Brokering: Substitutes dummy values in headers with real credentials without the agent ever seeing the secret.
  • Transparent Integration: Works with existing tools (MCP, CLI, SDKs) by bootstrapping the agent's environment via HTTPS_PROXY.
  • Egress Filtering: Allows operators to control which agents can access specific services and API endpoints.
  • Pluggable Stores: Can be backed by a local encrypted store or external secrets managers like Infisical for dynamic secrets.
  • Request Logging: Provides visibility into authenticated traffic to monitor and diagnose agent behavior.

Sources