AI Agent Security Checklist: Prompt Injection and Tool Access

By Numan ·

A practical AI agent security checklist covering prompt injection, MCP tool permissions, human approval, argument validation and failure testing.

AI agent security checklist: prompt injection and tool access

An AI agent becomes more useful when it can read files, search the web, and call tools. Those connections also create paths from untrusted content to real actions. This checklist helps a developer review an agent before giving it access to a service, repository, or workspace.

Start with the trust boundary: a web page, document, email, search result, or tool response may contain instructions, but it is still data. OWASP defines indirect prompt injection (https://genai.owasp.org/llmrisk/llm01-prompt-injection/) as behavior-changing input that arrives through an external source such as a website or file.

Inventory every tool and connection

Write down what the agent can read, change, send, and delete. Include local tools, MCP servers, browser access, and service accounts. For each connection, record the owner, permission scope, and the task that needs it. Remove unused connections.

This is a real development concern, not just a theoretical one. In a 2026 scan of nearly 10,000 developer environments (https://snyk.io/blog/agentic-development-security-ai-coding-risk/), Snyk reported security findings in some MCP setups. The sample and scanning method matter, so do not treat its rates as a universal measure; use the finding as a reason to inspect your own configuration.

Give each task the smallest useful permission

A summarizer usually needs read access. A draft creator may need permission to create drafts, but not publish or delete. Separate those capabilities in the underlying service. Keep credentials in the service or tool layer rather than placing them in prompts or retrieved text.

OWASP recommends least privilege (https://genai.owasp.org/llmrisk/llm01-prompt-injection/) because limiting access can reduce the impact of a successful injection. A friendly system prompt alone cannot enforce a permission boundary.

Treat retrieved content as untrusted