Prompt injection

Prompt injection is an attack that embeds adversarial instructions in text an AI system processes, causing it to follow the attacker’s intent instead of the operator’s. Because models read instructions and data from the same channel, any text the system ingests is a potential command channel.

The root cause is architectural: language models have no hard boundary between "content to analyze" and "instructions to follow." Injection is therefore managed, not patched. For agents the stakes escalate, because a successful injection does not just change words, it changes actions. Defenses layer input screening, output verification, and, decisively, governance of the actions an injected agent can actually take: an injection that cannot move a tool is a contained one.

Related

• Indirect prompt injection
• The 17 agentic attack techniques
• All glossary terms