Detect model output passed to dynamic execution or shell sinks

[Synvoya]

Problem

Model output is untrusted. Passing it into eval, Function, shell commands, or dynamically selected tool calls can turn prompt injection into code execution. CodeInspectus does not currently have a dedicated model-output-to-execution rule.

Scope

Start with high-signal, intrafile JavaScript/TypeScript flows from the model-output sources already recognized by the prompt-injection and dangerous-HTML analyzers into:

• eval(value)
• new Function(value) / Function(value)
• child_process.exec(value) and equivalent shell-string APIs

Unsafe tool-call dispatch may be proposed separately if it cannot meet the same precision bar.

Acceptance criteria

• Freeze a corpus with true positives and near misses before implementing the rule.
• Detect direct and split-variable intrafile flows from recognized model SDK output.
• Stay silent for constant strings, allow-listed dispatch, non-shell argument arrays, and validated/sanitized values where support is explicit.
• Emit a new ci-ai-* rule ID with remediation and OWASP/CWE mappings justified from primary sources.
• Document supported sources, sinks, and false negatives.
• npm run build && npm test && npm run eval pass.