MCP ยท Security Guide
Cybersecurity For AI Agents¶
AI agents now reach company files, databases and internal tools through the Model Context Protocol, known as MCP. Its maintainers state in the specification that MCP cannot enforce its own security principles at the protocol level. Whoever deploys it carries that duty.
Start With How MCP Works Read The Real Cases
New here? Take the 30-second tour
The Problem¶
What Changed - Through MCP, a language model can query databases, call APIs and run computations on systems outside the chat window. The specification describes tools as model-controlled, so the model finds and invokes them on its own from the user's prompt.
What Breaks - Each server supplies the text that describes its tools, and the model reads that text when it chooses what to do. Some clients showed a tool's description once at approval and gave no notice when it later changed, as Invariant Labs reported in April 2025.
Real Example - On 10 August 2026, one GitHub account opened 23 pull requests in 74 minutes to spread an MCP server named productivity-suite. The server formatted and summarised text as promised. After a client's third tool call, its tool descriptions changed and directed the connected agent to seek SSH keys and AWS credentials, and to hide the activity from the user. Pillar Security disclosed the campaign and found none of the pull requests merged when it reviewed them.
Start Where Your Problem Is¶
Checked Against Primary Sources
Protocol statements were checked against MCP specification revision 2026-07-28 on 12 November 2026. Incident details come from the researchers who first reported them, and every page links its sources.