Struct RegexSecurityAnalyzer

Source

pub struct RegexSecurityAnalyzer { /* private fields */ }

Expand description

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

Detects:

System prompt override attempts (“ignore previous instructions”, etc.)
Role injection (“system:”, “assistant:” in user messages)
Encoding attacks (base64-encoded malicious instructions)
PII patterns (email, phone, SSN, credit card)
Data leakage (system prompt leaks, credential exposure in responses)

§Example

use llmtrace_security::RegexSecurityAnalyzer;
use llmtrace_core::SecurityAnalyzer;

let analyzer = RegexSecurityAnalyzer::new().unwrap();
assert_eq!(analyzer.name(), "RegexSecurityAnalyzer");

Implementations§

Source §

impl RegexSecurityAnalyzer

Source

pub fn new() -> Result<Self>

Create a new regex-based security analyzer with all detection patterns compiled.

§Errors

Returns an error if any regex pattern fails to compile.

Source

pub fn with_jailbreak_config(jailbreak_config: JailbreakConfig) -> Result<Self>

Create a new regex-based security analyzer with custom jailbreak configuration.

§Errors

Returns an error if any regex pattern fails to compile.

Source

pub fn detect_injection_patterns(&self, text: &str) -> Vec<SecurityFinding>

Scan text against all injection patterns (including base64) and return findings.

This is exposed publicly so that the streaming security monitor can call it synchronously on content deltas without the async overhead of the full SecurityAnalyzer trait.

Source

pub fn detect_context_flooding(&self, text: &str) -> Vec<SecurityFinding>

Detect context window flooding attacks (OWASP LLM10: Unbounded Consumption).

Context window flooding is a Denial-of-Service technique where an attacker fills the LLM context window with junk content to crowd out legitimate instructions or inflate token-based costs.

Runs five heuristic checks:

Excessive input length — inputs exceeding 100,000 characters
High repetition ratio — >60% repeated word 3-grams
Low Shannon entropy — <2.0 bits/char on texts >5,000 characters
Invisible character flooding — >30% whitespace/invisible characters
Repeated line flooding — any single line appearing >20 times

This is exposed publicly so that the streaming security monitor can call it synchronously on content without the async SecurityAnalyzer trait.

Source

pub fn detect_pii_patterns(&self, text: &str) -> Vec<SecurityFinding>

Scan text for PII patterns and return findings.

Applies context-aware false-positive suppression: matches inside fenced code blocks, URLs, or well-known placeholder values are silently ignored.

Exposed publicly for use by the streaming security monitor.

Source

pub fn redact_pii( &self, text: &str, action: PiiAction, ) -> (String, Vec<SecurityFinding>)

Detect PII and optionally redact it from the text.

Behaviour depends on action:

Action	Returned text	Returned findings
`AlertOnly`	Original (unchanged)	All non-false-positive PII findings
`AlertAndRedact`	Redacted (`[PII:TYPE]`)	All non-false-positive PII findings
`RedactSilent`	Redacted (`[PII:TYPE]`)	Empty