# Lethal trifecta (AI agents)

**Definition:** The *lethal trifecta* (term popularized by Simon Willison) is the combination of **(1) untrusted input**, **(2) access to sensitive data or privileges**, and **(3) a channel outward** (links, images, APIs, tickets, email). When all three are present, a single successful injection can cause real damage.

**Why it matters:** many internal assistants look “safe” because only employees chat with them — yet they still index external text, see secrets, and render rich answers.

**Synthetic example:** untrusted connector error text (1) is indexed; the agent can read connector configs (2); the UI auto-loads markdown images (3). An injected instruction can place a secret into an image URL the browser fetches.

**What to do:** remove at least one leg — usually outbound rendering or excess privilege — then re-test. See [AI Security Review](/ai-security) and [indirect prompt injection](/guides/indirect-prompt-injection).

**Next:** [Cross-tenant retrieval](/guides/cross-tenant-retrieval) · [OWASP LLM Top 10 2026](/guides/owasp-llm-top-10-2026) · [Request a security review](/ai-security#scope)
