Prompt injection

Attack where adversarial input redirects the model away from its intended instructions. Mitigated by privileged-source system prompts and structured-input boundaries.

See also: allowlist