Skip to content
Search
SecurityPlaybooks

AI Agent Security Checklist: The Lethal Trifecta

Difficulty
Beginner
Time
30 minutes per agent

Before you start

  • The AI agent, assistant or automation you want to turn on, and access to its settings or a list of what it's connected to
  • Whoever set it up or manages it
  • A blank page or spreadsheet with three columns

This AI agent security checklist takes about thirty minutes and should happen before any agent is connected to your company's data. An agent is an AI tool that can take actions, such as reading files, searching the web or sending messages, through tool use or connectors such as MCP. Developer and writer Simon Willison named a dangerous combination the lethal trifecta: access to your private data, exposure to untrusted content, and the ability to communicate externally. When an agent has all three, text hidden in an email or web page it reads can steer it into sending your data out. This is called indirect prompt injection, and nobody yet knows how to prevent it reliably with filters or instructions. The dependable fix is to make sure no single agent has all three.

Step 1: List every private data source

In the first column, write everything the agent can read that isn't public: mailboxes, shared drives, documents, the CRM, calendars, chat channels, databases, customer records. Include what it can reach through other connected tools, not just what you meant it to use.

Step 2: List every way untrusted content reaches it

In the second column, write every source of text your company doesn't control: incoming email, web pages it browses or searches, uploaded files, support tickets, form submissions, shared documents from outside, chat messages from customers. If anyone outside your company can put words in front of it, that's untrusted content.

Step 3: List every way it can send information out

In the third column, write every outbound channel: sending email or messages, posting, creating public links, calling web addresses, submitting forms, writing to shared files others outside can see. Loading an image or a link can also count, because the address itself can carry data.

Step 4: If all three columns have entries, remove one

Pick the column you can live without for this job and remove it in the agent's settings, not with an instruction in its prompt. A line in a system prompt such as "never send customer data" can be talked around by injected text; a disabled permission can't. Examples:

  • An inbox assistant that reads email: remove its ability to send. It drafts; a person sends.
  • A research agent that browses the web: don't connect it to private files.
  • An agent that works on internal files and can message people: don't let it read outside email or the web.

If the job truly needs all three, it isn't ready to run unattended. Require a person to approve each outbound action, and get a security review first.

Step 5: Write it down and recheck when anything changes

Save the page with the agent's name, the owner, the date and what was removed. Recheck whenever someone adds a connector, plugin or permission, because one new tool can complete the trifecta.

Further reading

For the wider set of agent risks, the OWASP GenAI Security Project published the OWASP Top 10 for Agentic Applications in December 2025. Our glossary covers the related OWASP Top 10 for LLM applications and guardrails.

What to do next

Add the rule "no agent goes live without this checklist" to your team's AI use policy. Keep the completed pages with the policy so the next review starts from them.

Rather have someone set it up for you?

That's a fair call. LTFI, the agentic business operating system from Kief Studio, can set up the work these guides describe and run it with you. No website yet? Kief Studio will build you one for free.

Go further on kief.studio