The lethal trifecta is an AI security concept introduced by British developer Simon Willison on his blog in 2025. It holds that when an agent built on a large language model simultaneously has (1) access to private data, (2) exposure to untrusted content, and (3) the ability to communicate externally or take outside actions, an attacker can exfiltrate confidential data through prompt injection, an attack that hides instructions in content the model reads.
The key insight is that removing any one of the three elements breaks the attack path. Because models themselves cannot fully prevent prompt injection, the concept is widely cited as grounds for structurally separating permissions and data flows in agent design. It is frequently discussed in environments where agents combine multiple tools through connection mechanisms such as the Model Context Protocol (MCP).