Prompt Injection (LLM01:2025)

Manipulating an LLM’s input prompt to force the model to behave in an unintended manner. Ranked #1 in the OWASP LLM Top 10 (2025). Also a named risk in Google’s SAIF.

What It Is

Prompt injection exploits the fact that the prompt is the LLM’s only text-based input. By crafting or injecting malicious instructions, an attacker can:

  • Override the model’s system instructions or intended behavior
  • Bypass safety guardrails
  • Extract sensitive information
  • Cause the model to perform actions outside its intended scope

Relationship to Prompt Engineering

Prompt injection is the offensive application of prompt engineering. The same understanding of how phrasing, context, and constraints steer LLM output is used to craft malicious inputs that redirect the model.

OWASP Classification

LLM01:2025 — Prompt Injection covers security vulnerabilities that arise from manipulating an LLM’s input prompt, including:

  • Forcing the LLM to ignore its system prompt
  • Indirect injection via external content the LLM reads (documents, web pages)
  • Jailbreaking — overriding content policy restrictions

LLM02:2025 — Sensitive Information Disclosure — leakage of sensitive information resulting from improper prompt engineering or manipulation of the input prompt. Often achieved via prompt injection.

⚠️ Placeholder: attack techniques section not yet ingested.

See Also