Direct
Direct prompt injection is like a trickster sending a fake instruction to a digital assistant, pretending to be the owner to make it do things it shouldn't, such as revealing secrets or creating harmful content.
Direct prompt injection occurs when an attacker inputs malicious instructions directly into an LLM's prompt interface, bypassing intended constraints to manipulate the model's output, gain unauthorized system access, or force the generation of prohibited content.
Direct prompt injection is a security vulnerability where an adversary provides adversarial input directly to an LLM's prompt interface, effectively hijacking the model's instruction-following behavior. This exploit leverages the model's inability to distinguish between system instructions and untrusted user input, enabling the adversary to achieve unauthorized privilege escalation, exfiltrate sensitive data, or bypass safety guardrails to generate harmful content.