| name | prompt-injection |
| description | Detects LLM prompts built from user input or retrieved documents that let attackers hijack the model's instructions. Use when writing code that constructs LLM prompts from user input, builds system prompts, implements RAG pipelines, or processes external documents fed to a model. Also invoke when external data can influence LLM context. |
Prompt Injection Security Check (OWASP LLM01:2025)
What this checks
Protects against attacker-controlled text that hijacks LLM instructions. Direct
injection arrives through user input; indirect injection arrives through retrieved
documents, emails, or tool outputs. Both can cause the model to exfiltrate data,
bypass guardrails, or execute unintended actions.
Vulnerable patterns
- User input interpolated directly into the system-role message — user text lands in the instruction tier.
- Retrieved documents concatenated raw into the prompt with no delimiter or trust label.
- Email bodies, fetched web pages, or other external content passed into the prompt with no boundary markers separating data from instructions.
- No structural separation between developer instructions and untrusted data — everything is one string.
- Raw model response returned to the caller, rendered, logged, or used to trigger a downstream action with no validation step in between.
Fix immediately
Flag the vulnerable code and explain the risk. Then suggest a fix that establishes
these properties. Translate each property into the audited file's language and LLM
client library — use that library's documented role-separated message API rather
than mirroring an example from another stack.
- Trust tiers are structurally separate. Developer instructions go in the
system role; user input and retrieved documents go in the user role, wrapped
in explicit delimiter tags that label the content as untrusted data. Never
interpolate user text into the system prompt.
- Input is bounded and screened before the API call. Apply a length cap
and reject obvious injection markers (phrases like "ignore previous", "new
instruction"). Screening is a denylist and will not catch everything, but it
raises the bar.
- Output is validated before any downstream action. Every code path that
uses the model's response — returning it to the caller, rendering it,
logging it, triggering a tool call — first routes it through a gate that
enforces size bounds and rejects suspicious instruction language. A defined
validator that is never called does not satisfy this.
Verification
Confirm these properties hold:
References
