Security Policy¶

Estimated time to read: 1 minute

Supported Versions¶

LLMTrace is pre-1.0. Only the latest commit on main receives security fixes.

Reporting a Vulnerability¶

Do not open a public issue for security vulnerabilities.

Instead, please report them privately:

• Email: [email protected]
• GitHub: Use GitHub's private vulnerability reporting

Include: - Description of the vulnerability - Steps to reproduce - Affected component(s) (proxy, storage, security analysis, etc.) - Impact assessment (if known)

Response Timeline¶

We'll keep you informed throughout the process. If the issue is accepted, you'll be credited in the advisory (unless you prefer to remain anonymous).

Security Architecture¶

LLMTrace is designed with security as a core concern:

• Prompt injection detection — Regex-based and ML-based (DeBERTa) analysis of all proxied requests/responses
• PII detection & redaction — International PII patterns (UK NIN, IBAN, EU passports, NHS, etc.) with configurable redaction
• NER-based PII — BERT-based named entity recognition for detecting names, organisations, and locations
• Streaming security analysis — Real-time incremental checks during SSE streaming
• Statistical anomaly detection — Per-tenant sliding window analysis with configurable sigma thresholds
• OWASP LLM Top 10 — Built-in test coverage for all ten categories
• Per-tenant rate limiting — Token bucket rate limiting with Redis-backed isolation
• Cost caps — Per-agent budget enforcement with soft/hard limits
• RBAC — Role-based access control for tenant management
• Secrets management — No plaintext secrets in configuration; supports External Secrets Operator and Sealed Secrets

See the custom policies guide for configuration details.

Dependencies¶

We use Dependabot for automated dependency monitoring. Security-critical updates are prioritised.

Auditing¶

# Rust dependencies
cargo audit

# Container image
trivy image llmtrace-proxy:latest

# npm (dashboard)
cd dashboard && npm audit

Disclosure Policy¶

• Vulnerabilities are disclosed after a fix is available
• CVE IDs are requested for significant issues
• Advisories are published via GitHub Security Advisories