Skip to main content
Trident gives you end-to-end security coverage for AI agents — from the first prompt your agent receives to the findings that land in your inbox. This section covers the four pillars of agent security in Trident: distributed tracing to capture every LLM call and tool invocation, automated red-teaming to discover vulnerabilities before attackers do, a runtime firewall to block prompt injection and jailbreaks in production, and a findings inbox to triage, replay, and resolve issues.

Tracing

Automatically capture every prompt, tool call, and LLM response your agent makes — with zero-config PII redaction built in.

Red-Teaming

Run 200+ adversarial attack vectors against your agents covering the full OWASP Agentic Top-10.

Firewall

Scan every prompt and output for injection attacks, jailbreaks, and data exfiltration attempts in real time.

Findings

Review, replay, and resolve security issues from red-team runs, the firewall, static analysis, monitors, and agent self-reports.

​
End-to-end workflow

Trident’s agent security features are designed to work together across the full development and production lifecycle.
1

Instrument

Install the Trident SDK and call trident.init() (TypeScript) or vouch_sdk.init() (Python) once at startup. The SDK wraps OpenLLMetry and auto-traces every LLM call in the process — no additional code required.
2

Red-team

Before you ship, run a red-team campaign from the dashboard or via the API. Trident fires 200+ attack vectors at your agent and scores every result against the OWASP Agentic Top-10 and AIVSS severity scale.
3

Firewall

In production, route prompts through the Trident gateway proxy or call trident.scan() directly before each LLM call. Every prompt passes through a two-stage firewall: your project’s custom deny rules fire first, then the LLM Guard ensemble.
4

Review findings

All issues — from red-team campaigns, the runtime firewall, static analysis, monitors, and agent self-reports — land in the Findings inbox. Confirm a finding to automatically push a new firewall rule. Replay it at any time to verify your fix.

​
Supported frameworks

Trident’s tracing and firewall integrations work with every major LLM framework and provider. Install the SDK and your existing code is instrumented automatically — no framework-specific plugins required. LLM providers: OpenAI, Anthropic, Amazon Bedrock, Google Vertex AI, Cohere Agent frameworks: LangChain, CrewAI, LlamaIndex Protocols: Model Context Protocol (MCP)

​
PII redaction

Trident redacts sensitive data in your process before any trace data leaves your infrastructure. The following patterns are scrubbed from span attributes automatically:
  • Email addresses
  • Credit card numbers (validated with Luhn checksum)
  • US Social Security Numbers (SSNs)
  • AWS access key IDs
  • JSON Web Tokens (JWTs)
  • API keys (OpenAI, Anthropic, and similar sk- prefixed secrets)
  • IBANs
  • IPv4 addresses
  • Phone numbers
You can supply your own redaction rules or disable redaction entirely by passing options to trident.init(). See the Tracing page for details.