Keeper lists five ways MCP servers leak credentials

Keeper's teardown names five ways Model Context Protocol servers leak credentials, per The Hacker News. MCP lets an agent reach a database, cloud account, or internal API through one config file. Anyone who gets that file inherits whatever access the agent holds, across every system it touches.

The first two vectors are storage failures. Credentials sit in plaintext inside the server's config file, and no central inventory tracks which server holds which secret or when it was rotated.

The third is behavioral: prompt injection can manipulate an agent into misusing a credential it already holds legitimately.

The fourth and fifth are permissions and provenance. Access scoped for local development often carries into production without review. Untrusted servers can execute code on install too, a risk the piece ties to CVE-2025-6514.

Keeper wrote the piece itself, so it reads as prescriptive advice, not independent research.

Inventory and rotation are standard practice for any other credential system. MCP tooling doesn't make either easy yet, so most teams will keep skipping it until an incident forces the audit.