Skip to content
All Posts
AI Integration

Lethal Trifecta and Rule of Two: Is Your AI Agent a Data-Leak Risk?

Liquid Intent|November 15, 2026|7 min read
Share

Most conversations about AI agent risk stall on abstractions: “the model might hallucinate,” “the vendor might train on our data.” Those are real concerns, but they do not help a leader decide a concrete question: is it safe to let this agent run without a person watching it?

There is a simple test for that question. It comes from security researcher Simon Willison, who calls it the “lethal trifecta,” and it pairs with a design rule from Meta called the “Agents Rule of Two.” Together they give business leaders a way to size up an agent before it touches anything that matters.

Why Prompt Injection Is the Problem to Plan Around

The OWASP GenAI Security Project's State of Agentic AI Security and Governance report (June 2026) identifies prompt injection as a leading security vulnerability in agentic systems, connecting it to six of the ten categories in its Top 10 for Agentic Applications.

Prompt injection means that text an agent reads can be treated as an instruction. An AI model does not reliably separate “content I am summarizing” from “orders I should follow.” If an email, web page, support ticket, or shared document contains hidden instructions, the agent may act on them as though your own team had written them.

That is why the question to ask is not “will the model behave?” It is “what could this agent do if someone fed it a malicious instruction?” The trifecta answers that.

The Three-Part Test

According to Willison’s “lethal trifecta,” an agent becomes a data-exfiltration vector when it combines three capabilities:

  1. Access to private data. The agent can read things that are not public: customer records, contracts, internal documents, email, financials, source code, credentials.
  2. Exposure to untrusted content. The agent reads material that someone outside your control could have written: inbound email, web pages, uploaded files, support tickets, third-party messages, public reviews.
  3. The ability to communicate externally. The agent can send information out: send an email, post a message, call an outside API, load a URL, or write to a shared location others can see.

Any one of these is manageable. The danger is the combination.

Why the combination leaks data

Walk through it as a sequence. An attacker cannot reach your private data directly. But they can put text where your agent will read it (capability two). That text tells the agent to gather what it has access to (capability one) and send it somewhere the attacker can collect it (capability three). The agent is not hacked in the traditional sense. It simply follows instructions, which is what it was built to do.

Illustrative scenario

Imagine an assistant that triages a shared inbox, can look up customer accounts, and can send replies. A message arrives containing hidden text telling the assistant to look up a customer's account details and include them in a reply to an outside address. All three capabilities are present, so nothing in the design stops it.

Removing any one capability interrupts this particular path to data theft. Without access to private data, untrusted instructions, or an external communication channel, this type of attack becomes significantly harder. That doesn’t eliminate every security risk, but it reduces the agent’s exposure to this specific threat.

The Rule of Two

Meta’s “Agents Rule of Two” builds on that observation with a broader design rule: an autonomous agent should have no more than two of three capabilities without additional safeguards. Meta’s framework also considers the ability to modify systems or data, extending the concern beyond information leaks to potentially harmful actions. If a task genuinely requires all three, the agent should not operate without additional safeguards, such as human approval or another reliable validation mechanism.

In plain terms, that gives you three safe-ish configurations and one that needs supervision:

  • Private data + untrusted content, no external communication. An agent that summarizes inbound documents against internal records, but can only return results to a person inside your organization.
  • Private data + external communication, no untrusted content. An agent that works only from trusted internal inputs and can send reports to known recipients.
  • Untrusted content + external communication, no private data. A public-facing agent that reads outside material and replies, but has no access to anything sensitive.
  • All three. Allowed only with a human approving the consequential steps, such as sending anything out.

The rule does not say agents must be weak. It says autonomy should be spent deliberately. Decide which two capabilities the task truly requires, and give the agent a human checkpoint for the third.

Putting the test to work

You can run this test in a meeting, with no technical background:

  1. List every agent or AI automation in the business, including ones teams set up on their own.
  2. For each, ask three yes/no questions: Can it read private data? Does it read content from outside the company? Can it send or publish anything outward?
  3. Any agent with three “yes” answers and no human approval step is the first one to fix.

That inventory step matters more than it sounds. IBM data cited in the OWASP report found that only 37% of organizations have policies to detect Shadow AI. If you cannot see which agents exist, you cannot apply the test to them. We cover that visibility gap in Ungoverned AI Is Becoming Shadow Operations.

What an Independent Security and Architecture Review Should Check

The trifecta is a screening tool, not a full assessment. When an agent combines the capabilities or you are unsure, an independent review, meaning someone who did not build the system, should answer these questions in language you can act on:

  • What can the agent actually reach? Not what it is meant to use, but what its credentials allow. Agents are often given broad access because it was easier during setup.
  • Where does untrusted content enter? Every inbox, form, file upload, web lookup, and third-party integration is a path for injected instructions.
  • What are the exits? Email, messages, outbound web requests, file sharing, and API calls. If the agent can send data out, where can it send it, and is that limited to known destinations?
  • Where do humans approve? Which actions need sign-off, and is the approval real, with enough context to judge, or a button people click without reading?
  • What is logged? Can you reconstruct what the agent read, decided, and sent after an incident?
  • Can you undo it? If the agent does the wrong thing, is there a rollback path? See AI Agent Rollbacks and Post-Launch Operations.
  • Who owns it? Someone should be named as accountable for the agent’s permissions and behavior over time.

This independent view matters most for systems that AI helped build. The reasons are laid out in The AI That Built Your System Should Not Be the Only One Checking It.

What to Do This Quarter

You do not need to halt AI adoption to act on this. A practical sequence:

  • Inventory your agents and automations, including unofficial ones.
  • Run the three-question test on each.
  • For any agent with all three capabilities, remove unnecessary access or introduce safeguards such as human approval, restricted external destinations, or independent validation before consequential actions.
  • Narrow permissions to what each task needs.
  • Have someone independent review the highest-risk systems.

Liquid Intent helps organizations design, integrate, and review AI systems so that autonomy is placed where it is safe. You can see what we do on our services page.

Not Sure Whether Your Agents Pass the Test?

Tell us what your AI agents can access, read, and send. We will help you work out where a human approval step belongs.

Talk to Us