Theory · 10 min · AI at work: what the SOC sees
Agents, tools and MCP
Explain how an AI agent turns an injection into an action on a real system.
An AI agent is a model that can call tools: run a shell command, read files, query a database, send an email. The Model Context Protocol (MCP), introduced by Anthropic in November 2024 and now widely supported, is a standard way to plug tool servers into assistants and agents.
Tools turn words into actions. If an injection can steer the agent, it can now steer the tools, with the agent’s permissions. OWASP calls the root cause Excessive Agency (LLM06): too much functionality, too many permissions, too much autonomy.
The OWASP Top 10 for Agentic Applications (December 2025) names the follow-on risks, starting with ASI01 Agent Goal Hijack and ASI02 Tool Misuse and Exploitation. A malicious or compromised MCP server can also lie in its own tool descriptions, which the model reads as instructions ("tool poisoning").
For the SOC the endpoint view is familiar: a process tree. The agent process starts node, PowerShell or curl. What is new is that the "user" behind those commands may be text the agent read in an issue, an error report or a web page.
T1698T1700AML.T0053AML.T0086AML.T0110
Your turn: select the field that proves it. A coding agent read an SSH private key. Select the field that shows where the instruction to do so came from.
Do the interactive check Free, in the browser.
Sources
- Model Context Protocol specification (security and trust) Model Context Protocol
- OWASP Top 10 for LLM Applications 2025 OWASP GenAI Security Project
- OWASP Top 10 for Agentic Applications for 2026 (December 2025) OWASP GenAI Security Project
- MCP security notification: tool poisoning attacks (April 2025) Invariant Labs
- ATT&CK T1700: AI-Enabled Exfiltration MITRE ATT&CK v20
- Research note: Agentjacking via MCP and Sentry injection (June 2026) Cloud Security Alliance