IonBin
All pages
Search
Concept

The lethal trifecta (agent security)

The lethal trifecta is Simon Willison’s name for combining private-data access, untrusted input and an outbound communication path in an AI agent. It highlights a data-leak risk. Removing a capability or enforcing strict boundaries can reduce exposure; other prompt-injection harms still need controls.

Verified Oct 18, 2026 Review by Apr 16, 2027 2 sources

What it is

Simon Willison’s June 2025 article names a combination that enables data theft through prompt injection: an agent can read sensitive material, encounter attacker-controlled content and transmit information outside its trusted boundary.

How it works

An email, document or web page can contain instructions that try to redirect the agent. If it follows them while holding private information and a usable outbound tool, it may disclose that information. Combining MCP servers can assemble these capabilities unintentionally.

Willison recommends avoiding that combination when mixing tools. This is a useful threat model, not a claim that every such session is successfully compromised.

Benefits

IonBin’s suggested review is to map each workflow’s readable data, untrusted inputs and possible destinations. Then remove unnecessary permissions and have an engineer enforce boundaries outside the prompt. Test the resulting boundary using harmless dummy secrets before connecting real records.

OWASP recommends least privilege, separation of external content, adversarial testing and human approval for high-risk actions. Use these as layered controls rather than treating a detection percentage as a security guarantee.

Limits and risks

The trifecta concerns exfiltration. Prompt injection can also distort answers, manipulate decisions or trigger unauthorized actions without all three legs.

A sandbox only helps with the resources and network paths it actually restricts. Inspect its configured permissions; the label alone says little about possible data transfers.

Review consequential writes, purchases and deletions as well as outbound messages. Keep approval prompts specific about the proposed action, destination and data so the reviewer can assess them. Splitting a workflow into agents does not create a security boundary if they retain the same permissions or pass untrusted instructions between them.

Best-fit stacks

Stack 08

Dev, web & IT

Design, code, host and ship sites and internal tools, with an AI coding agent doing most of the typing.

Stack 06

Ops & admin automation

Hand repeat admin (receipts, logs, notices, handoffs) to workflows your agent can trigger.

Agent compatibility

Tool in this entryHow agents connect
Claude Agent platform MCPAPI/CLIFile import
GitHub MCPAPI/CLIFile import

Levels come from each tool's recorded, sourced compatibility. How the levels work.

Sources

  1. The lethal trifecta for AI agents: private data, untrusted content, and external communication · Simon Willison, 2025-06-16Practitioner view accessed Oct 18, 2026
  2. LLM01:2025 Prompt Injection · OWASP Gen AI Security Projectofficial accessed Oct 18, 2026

Practitioner views are restated in our words and labeled as views, not measured results. Product facts come from official sources dated above.

Suggest an edit

IonBin's wiki isn't user-edited. Corrections go to our Directory Curator, who checks them against primary sources before anything changes. Spot something wrong or out of date on “The lethal trifecta (agent security)”? Email us with a link to a primary source.

Suggest an edit by email or write to contact@ionbin.com