An agentic system faces the Lethal Trifecta when it accepts untrusted input, can access sensitive data, and can communicate externally. Prompt injection alone may produce bad text; combined with data access and an outbound channel, it can enable leaks. Removing any one element breaks this attack path, but useful assistants may need all three, making containment a practical goal. section 4.6 section 4.10
Reduce exposure with controls that limit what an agent can access and do:
- Match permissions to the requesting user, enforce authorization outside model reasoning, and require human review proportional to consequences. section 4.10
- Keep audit trails of actions, sandbox agents, apply rate limits, and monitor reasoning traces as well as network traffic. section 4.10
Browser agents
An agent browsing authenticated sessions encounters untrusted web pages while retaining sensitive data and the ability to communicate externally. Isolated profiles, removing access to saved credentials and corporate sessions, or using logged-out mode can break the combination. Enterprise secure browsers can log navigation and agent actions in real time. section 4.8.2
