Skip to main content
TrustGuard ships a fixed catalog of built-in detectors. You don’t write detectors — you pick one from the catalog and configure it (see Detectors). This page is the index of what’s available; each detector’s full settings are documented on its category page (linked below). The console’s Detectors → Catalog tab lists the catalog grouped by category, with a configuration form for each detector’s settings.

​
Categories

CategoryFocus
Application securityClassic injection/code attack patterns in the request.
Data loss preventionPII and secret detection + masking.
Content securityJailbreaks, toxicity, moderation, document/URL analysis.
Behavioral securityAbusive-actor and anomaly detection across requests.
Agent & MCP securityTool/function definitions and tool-call validation.

​
The catalog

Each detector is identified by a stable slug. Sides = which directions it supports. Mutable detectors can rewrite the payload (so they support redact).

​
Application security

Detector (slug)DetectsSidesProtocolsMutable
code_sanitationDangerous code-injection patterns by language (JS, Python, PHP, SQL, shell, HTML) + custom patterns.inputall—
injection_protectionClassic injection patterns (SQL, NoSQL, command, path traversal, XSS, LDAP, XPath, header, file inclusion) in chosen request scopes.inputall—

​
Data loss prevention

Detector (slug)DetectsSidesProtocolsMutable
data_loss_prevention60+ PII entities and secrets (passwords, API keys, tokens, JWTs); masks them in flight.input, outputall✅

​
Content security

Detector (slug)DetectsSidesProtocolsMutable
prompt_guardJailbreaks / prompt injections, scored by the NeuralTrust Firewall.input, outputall—
multiturn_guardMulti-turn jailbreaks that build up across a session_id.inputall—
toxicityToxic content, via NeuralTrust Firewall or OpenAI.input, outputall—
toxicity_openaiToxicity via the OpenAI Moderation API.input, outputall—
toxicity_azureToxicity via Azure AI Content Safety.input, outputall—
prompt_moderationOff-topic / disallowed content via keyword+regex and/or topic probability.input, outputall—
url_analyzerFetches URLs in the content (SSRF-guarded) and screens them for jailbreaks and PII.inputllm, mcp—
doc_analyzerExtracts text from uploaded documents (incl. OCR) and screens for PII and jailbreaks.inputllm—
bedrock_guardrailApplies an AWS Bedrock guardrail (topic / content / sensitive-info policies).bothall—

​
Behavioral security

Detector (slug)DetectsSidesProtocolsMutable
anomaly_detectorAbusive actors keyed on consumer_id: bot-like timing, repeated payloads, escalation, cross-collector abuse.inputall—

​
Agent & MCP security

Detector (slug)DetectsSidesProtocolsMutable
tool_guardJailbreaks/injections planted in the agent’s own system prompt and tool descriptions.inputmcp—
tool_permissionTools requested in an MCP call against an allow/deny list.inputmcp—
tool_selectionTool calls the model emitted, against a known-tool catalog (hallucinated tools, bad arguments).outputmcp—
Hidden detectors. injection_protection, tool_permission, and tool_selection are functional but currently not shown in the catalog picker. Contact NeuralTrust if you need them enabled for your team.

​
How detectors and modes fit together

  • A catalog detector is a fixed capability — you can’t change its code, only its settings.
  • You create a named, reusable instance of it with a mode, protocol, direction, and settings, then attach it to collectors. See Detectors.
  • Only data_loss_prevention is mutable — it’s the only detector where redact is valid and the only one that can populate transformed_payload.
Next: configure a detector, attach it to a collector, and inspect traffic with the Guard API.