AIR-007high SeverityAI-Generated Anti-Patterns & Dynamic Execution
Prompt Injection Safeguards on User Input Before LLM Calls
Agents concatenate raw user input directly into system prompts (e.g. `System: ${userPrompt}`) without boundary delineation.
1-Click Fix Prompt for Cursor & Claude Code
Review my codebase for "Prompt Injection Safeguards on User Input Before LLM Calls" (AIR-007).
Issue: Agents concatenate raw user input directly into system prompts (e.g. `System: ${userPrompt}`) without boundary delineation.
Remediation required: Use structured role messages (role: "user", role: "system") and separate untrusted input from system instructions.
Please:
1. Locate any occurrences of this issue across my repository.
2. Refactor the code to follow production standards without breaking existing functionality.
3. Provide a clear command or step to test and verify the fix.Why AI Assistants Write This Bug
Agents concatenate raw user input directly into system prompts (e.g. `System: ${userPrompt}`) without boundary delineation.
AI code generation prioritizes immediate syntax execution over production edge cases. When an assistant hits an undefined variable or strict type error, it frequently takes the path of least resistance rather than implementing standard pre-launch security patterns.
Code Comparison
Common AI Code PatternVulnerable
// Vulnerable AI-generated pattern for: Prompt Injection Safeguards on User Input Before LLM Calls
// Problematic pattern: Agents concatenate raw user input directly into system prompts (e.g. `System: ${userPrompt}`) without boundary delineation.Launch-Ready Fix
// Production-ready fix for: Prompt Injection Safeguards on User Input Before LLM Calls
// Recommended remediation: Use structured role messages (role: "user", role: "system") and separate untrusted input from system instructions.Remediation & Verification Checklist
Remediation: Use structured role messages (role: "user", role: "system") and separate untrusted input from system instructions.
- Search your codebase for pattern matching "Prompt Injection Safeguards on User Input Before LLM Calls".
- Apply the recommended remediation: Use structured role messages (role: "user", role: "system") and separate untrusted input from system instructions.
- Run local build and test commands to ensure no regressions.
- Verify in your browser or with curl that the issue is resolved.
Audit All 108 Checks in One Click
Checking items one by one takes hours. Let us do it in 2 minutes.
With the Founder Pass ($15 one-time), we automatically scan your repository for this issue, plus all other 107 launch checks across secrets, Stripe webhooks, Supabase RLS, and security headers.