User input interpolated into system prompt
User-controlled data is interpolated directly into a system prompt. Users can override your instructions ('ignore previous instructions…'), exfiltrate the prompt, or steer the model into unintended behavior.
What the check contributes
Rule LLM002runs as part of CodeCordon's deterministic static-analysis pass. When it matches, the finding includes the source file, line number, severity, explanation, and a redacted snippet so the developer can inspect the evidence directly.
Recommended remediation
Keep system prompts static. Pass user data in user-role messages, delimit it clearly, and treat model output that acts on it as untrusted.
How to use this result
- 1. Open the reported line. Confirm the match is active application code rather than a fixture, example, or false positive.
- 2. Apply the remediation in context. Preserve the intended behavior while removing the dangerous pattern.
- 3. Scan again. Verify the known pattern no longer appears, then continue with tests and any runtime or human security review appropriate to the application.
Limits of this check
A match is a review signal, not proof of exploitability. No match means only that this specific known pattern was not found in the applicable files. It does not rule out equivalent code, business-logic flaws, runtime vulnerabilities, or novel attack paths.