Prompt Injection
Definition
Prompt injection is input that causes an AI application to ignore its instructions, often to reach tools, data or actions it should not.
What Is Prompt Injection?
Prompt injection exploits the fact that LLM applications mix instructions and data in the same text. An attacker crafts text that the model treats as new instructions.
Types of Prompt Injection
- Direct: the attacker types the malicious instruction
- Indirect: the instruction hides in content the AI reads, such as a web page, email or document
Prompt Injection Examples
- A support bot persuaded to reveal its system prompt
- An email that instructs an assistant to forward the inbox
- A document that makes an agent call an internal API
Risks of Prompt Injection
Impact grows with what the AI can access. An injected chatbot embarrasses; an injected agent with tool access can leak data or take actions. AI red teaming tests these paths end to end.
Related ranking
Related terms
Bug Bounty ProgramA bug bounty program pays independent security researchers for valid vulnerabilities they report in an organization's in-scope assets.External Attack Surface ManagementExternal attack surface management (EASM) is the continuous discovery and monitoring of an organization's internet-facing assets and their exposures.Breach and Attack SimulationBreach and attack simulation (BAS) runs simulated attacker techniques against an environment to test whether security controls detect and block them.API Penetration TestingAPI penetration testing is the security testing of application programming interfaces by attempting real attacks against their authentication, authorization and business logic.