The Classic Paradox: Trusting AI Agents with Hardcoded Secrets
Whenever teams start integrating autonomous AI agents into their core infrastructure (GitLab, MariaDB, OpenProject, Portainer), the first instinct is often: "Let's just pass the API token or DB password into the system prompt or environment rules so it can get things done."
As a long-time sysadmin and architect, this raises immediate red flags. An AI agent is a non-deterministic process. It can hallucinate, suffer from prompt injections, or have its context transcripts logged. Handing it raw credentials isn't delegation—it's a scheduled security breach.
In my latest technical pattern, I wanted to solve a straightforward challenge: How do we let an AI agent interact with internal systems without ever exposing a single password, PAT, or SSH key to the model?
The Core Approach: Inverting Secret Ownership
Instead of giving secrets to the agent, we shift secret ownership entirely to a Model Context Protocol (MCP) server backed by HashiCorp Vault:
Logical Handles Only: The agent only manipulates short identifiers (resource_id). It never sees, holds, or transmits passwords or tokens.
Isolated MCP Runtime: The MCP container authenticates to Vault via AppRole, fetches credentials in-memory, opens backend sessions, and returns only the operational result to the agent.
Scoped Action Policies: Tools are restricted at the Vault record level (actions allowlist with MCP_ACTION_POLICY=enforce). Even if a tool exists in the container, Vault denies execution if it's not authorized for that specific resource.
Human-in-the-Loop Control: Mutating actions require explicit confirm: true parameters enforced by host rules.
Trajectory to V2 (LDAP Identity): Moving from container-wide AppRoles to propagating session identity down to Vault ACLs to eliminate the "confused deputy" flaw.
Let's Discuss: How Are You Handling Secret Governance?
I've published the full technical blueprint—including Mermaid sequence flows, Vault KV schemas, and sample MCP payloads—on GitLab.
I’d love to hear how other engineers and architects on dev.to are tackling these boundaries:
How do you currently isolate AI agents from core infrastructure secrets?
Are you relying on native MCP servers, proxy layers, or custom API wrappers?
What mechanisms do you use to enforce human validation before an agent mutates state?
Looking forward to reading your feedback, critiques, and alternative approaches in the comments!
🇬🇧 English version:
https://gitlab.com/Aztechnology/articles/-/blob/main/no-password-for-my-agent.md
🇫🇷 Version française :
https://gitlab.com/Aztechnology/articles/-/blob/main/agent-sans-mot-de-passe.md