Prompt Injection

Prompt injection is when an attacker mixes instructions disguised as system ones into the context via external content (web pages, emails, documents), hijacking the agent’s behavior. It’s more dangerous in agentic systems than in chatbots: an agent can perform irreversible actions. Context-level protection: marking the source of external content, strict message roles (separating instructions and data), input sanitization. But this is only the first line of defense – execution-level protection is needed.

Related: [Fences], [Deadly Triad], [Agent Loyalty]