Model output executed as code or shell command
LLM output is passed to eval, exec, or a shell. A prompt-injected model (or just a hallucinating one) gets arbitrary code execution on your server.
What the check contributes
Rule LLM004runs 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
Never execute raw model output. Use structured tool-calling with a fixed set of functions, validate arguments against schemas, and sandbox anything that must run generated code.
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.