Blog · MCP & API security

Notes from thegoverned gate.

Real incidents in MCP and API security — and what they teach us about giving AI access to the software you already run, without giving away the keys.

Prompt Injection ShadowLeak: the agent read an e-mail nobody could see One invisible instruction in an e-mail. The agent obeyed it with everything its owner had connected. Read it → MCP Security The state of MCP security going into 2026 Tool poisoning, prompt injection, over-scoped tokens: the attack classes have settled. Read it → Prompt Injection The GitHub MCP data heist: when one issue drains a private repo One malicious issue in a public repo, and the agent leaked the private one next door. Read it → API Security The raw endpoint problem: why unrestricted APIs keep leaking The same bug, over and over: an endpoint that answers more than the caller should get. Read it → Token Security One token, 700 companies: the Salesloft Drift OAuth breach A single stolen integration token opened the CRM of roughly 700 companies at once. Read it →