Prompt Injection

Prompt Injection is an application-level security vulnerability and a form of Adversarial Attack affecting Large Language Model applications and AI Agent workflows, in which attacker-controlled content attempts to override trusted instructions or redirect system behavior. It can occur directly through crafted user input or indirectly through external content, such as web pages, files, or tool outputs, that the application passes to the model. The underlying problem is inadequate separation of trusted instructions from untrusted data, allowing embedded instructions to be treated as authoritative. Successful attacks may circumvent Safety Guardrails, override directives in a System Prompt, cause Data Exfiltration, generate restricted content, or trigger unauthorized tool use through Function Calling. Organizations including OpenAI, Anthropic, and Google publish mitigation research and guidance emphasizing layered defenses, such as Input Sanitization, attack-detection Classifiers, restricted tool permissions, and user confirmation for sensitive actions; these measures reduce risk but do not guarantee prevention.