Skip to main content
Application-security detectors catch traditional web-attack patterns that ride into AI systems through prompts, tool arguments, and request bodies — SQL/command injection, code execution, XSS, path traversal, and similar. They run in-process (no external calls) and are detection-only (observe or block).
DetectorSlugSidesProtocolsModes
Injection Protectioninjection_protectioninputallobserve, block
Code Sanitationcode_sanitationinputallobserve, block
injection_protection is functional but currently not shown in the catalog picker. Contact NeuralTrust if you need it enabled for your team.

​
Injection Protection

Scans the request scopes you select for known injection-attack signatures and reports each match as a finding. Detected attack types: sql, nosql, command, path_traversal, xss, ldap, xpath, header, file_inclusion (or all).

​
Settings

FieldTypeRequiredNotes
content_to_checkarray<enum>✅Scopes to scan: headers, path_and_query, body, all.
predefined_injectionsarray<{ type, enabled }>—Built-in attack types to enable. Empty = all enabled.
custom_injectionsarray<{ name, pattern, content_to_check }>—Your own regex signatures, scoped per scope.
{
  "name": "Block injections (body)",
  "type": "injection_protection",
  "mode": "block",
  "direction": "input",
  "settings": {
    "content_to_check": ["body"],
    "predefined_injections": [
      { "type": "sql", "enabled": true },
      { "type": "command", "enabled": true }
    ],
    "custom_injections": [
      { "name": "ssti", "pattern": "\\{\\{.*\\}\\}", "content_to_check": "body" }
    ]
  }
}

​
Code Sanitation

Detects dangerous code-injection patterns by language in the request. Reports matches; never rewrites. Built-in languages: javascript, python, php, sql, shell, html.

​
Settings

FieldTypeDefaultNotes
apply_all_languagesbooleanfalseEnable every built-in language pattern set.
languagesarray<{ language, enabled }>—Enable specific languages.
custom_patternsarray<{ name, pattern, description, content_type }>—Custom regex signatures.
{
  "name": "Flag code injection",
  "type": "code_sanitation",
  "mode": "observe",
  "direction": "input",
  "settings": {
    "apply_all_languages": false,
    "languages": [
      { "language": "shell", "enabled": true },
      { "language": "python", "enabled": true }
    ]
  }
}

​
When to use

  • Put these in front of any system that forwards prompt/tool content into databases, shells, templating engines, or downstream APIs.
  • Start in observe to baseline false positives on your real traffic, then promote to block.
  • These are pattern-based and complementary to the LLM-aware content-security detectors — run both.