Crate llmtrace_security 

Security analysis engines for LLMTrace

This crate provides regex-based security analyzers for detecting prompt injection attacks, encoding-based attacks, role injection, PII leakage, and data leakage in LLM interactions.

Feature: ml

When the ml feature is enabled, an ML-based analyzer using the Candle framework becomes available:

• [MLSecurityAnalyzer] — runs local inference with a HuggingFace text classification model (BERT or DeBERTa v2).
• [EnsembleSecurityAnalyzer] — combines regex and ML results for higher accuracy.

Re-exports

pub use jailbreak_detector::JailbreakConfig;

pub use jailbreak_detector::JailbreakDetector;

pub use jailbreak_detector::JailbreakResult;

pub use action_policy::ActionPolicy;

pub use action_policy::ContextMinimizer;

pub use action_policy::EnforcementMode;

pub use action_policy::Message;

pub use action_policy::PolicyDecision;

pub use action_policy::PolicyEngine;

pub use action_policy::PolicyVerdict;

pub use canary::CanaryConfig;

pub use canary::CanaryDetection;

pub use canary::CanaryToken;

pub use canary::CanaryTokenStore;

pub use tool_firewall::FirewallAction;

pub use tool_firewall::FirewallResult;

pub use tool_firewall::FormatConstraint;

pub use tool_firewall::FormatViolation;

pub use tool_firewall::MinimizeResult;

pub use tool_firewall::SanitizeDetection;

pub use tool_firewall::SanitizeResult;

pub use tool_firewall::StrippedItem;

pub use tool_firewall::ToolContext;

pub use tool_firewall::ToolFirewall;

pub use tool_firewall::ToolInputMinimizer;

pub use tool_firewall::ToolOutputSanitizer;

pub use tool_registry::ActionRateLimiter;

pub use tool_registry::RateLimitExceeded;

pub use tool_registry::ToolCategory;

pub use tool_registry::ToolDefinition;

pub use tool_registry::ToolRegistry;

pub use action_correlator::ActionCorrelator;

pub use action_correlator::CorrelationConfig;

pub use action_correlator::CorrelationResult;

pub use action_correlator::TrackedAction;

pub use adversarial_defense::AdversarialDefense;

pub use adversarial_defense::AdversarialDefenseConfig;

pub use adversarial_defense::MultiPassNormalizer;

pub use adversarial_defense::PerturbationDetector;

pub use fpr_monitor::FprDriftAlert;

pub use fpr_monitor::FprMonitor;

pub use fpr_monitor::FprMonitorConfig;

pub use mcp_monitor::McpMonitor;

pub use mcp_monitor::McpMonitorConfig;

pub use mcp_monitor::McpSecurityViolation;

pub use multi_agent::AgentId;

pub use multi_agent::AgentProfile;

pub use multi_agent::MultiAgentConfig;

pub use multi_agent::MultiAgentDefensePipeline;

pub use multi_agent::TrustLevel;

pub use result_parser::AggregatedResult;

pub use result_parser::AggregationStrategy;

pub use result_parser::DetectorResult;

pub use result_parser::DetectorType;

pub use result_parser::ResultAggregator;

pub use result_parser::ScanResult;

pub use result_parser::ThreatCategory;

pub use session_analyzer::SessionAnalysisResult;

pub use session_analyzer::SessionAnalyzer;

pub use session_analyzer::SessionAnalyzerConfig;

Modules

action_correlator

Multi-step action correlation for cross-request attack sequence detection.

action_policy

Action-selector policy enforcement and context minimization.

adversarial_defense

Adversarial ML robustness module (R-IS-08).

canary

Canary token system for detecting system prompt leakage (OWASP LLM07).

code_security

Code security analysis for LLM-generated code outputs.

fpr_monitor

Production FPR monitoring with drift detection (R-IS-01).

jailbreak_detector

Dedicated jailbreak detection module.

mcp_monitor

MCP Protocol Security Monitoring (R-AS-06).

multi_agent

Multi-Agent Defense Pipeline (R-AS-04).

normalise

Unicode normalisation layer for security analysis.

pii_validation

PII checksum validation to reduce false positives.

result_parser

Standardized tool result parsing (R-AS-01).

session_analyzer

Multi-turn session analysis for detecting extraction attacks (R-IS-03).

tool_firewall

Tool-boundary firewalling for agent security.

tool_registry

Tool registry and action-type rate limiting for agent security.

Structs

RegexSecurityAnalyzer

Regex-based security analyzer for LLM request and response content.

Functions

is_likely_false_positive

Check whether a PII match is likely a false positive based on its context.