Everything on Threat Modeling

The 'Lethal Trifactor' identifies that prompt injection is only dangerous when an agent has access to private data, external communication capabilities, and untrusted input. Breaking any one of these links neutralizes the threat.

Impact: Security teams can prioritize controls by focusing on isolating data or restricting network egress, rather than relying on imperfect model-level filters.

— from Sandboxing AI Agents to Mitigate Prompt Injection Risks · INNOQ Podcast· Mar 23, 2026