Blog
Notes from building the runtime.
Detection research, engineering decisions we had to defend, and threats we found while building. Written by the people who wrote the code.
LLM as Judge: Making the Calls Rules Cannot
LLM-as-judge for guardrails: which calls need semantic judgement, why judges ship advisory-only, multi-judge consensus, calibration, and boundary tradeoffs.
Rules vs. Models: Layering Deterministic and ML Detection
When to escalate from deterministic rules to an ML classifier: uncertainty bands, latency budgets, graceful fallback, and keeping inference in your boundary.
What Are AI Guardrails? A Practical Guide for Production LLM Apps
A practical guide to AI guardrails for production LLM apps: what they are, prompt injection and PII detection, where checks run, and how to adopt them.
Prompt Injection: How It Works and How to Defend Against It
A practical guide to prompt injection and indirect prompt injection: how the attack works, why system prompts fail, and layered LLM security defenses.
Preventing Data Leaks: PII and Secret Detection in LLM Pipelines
PII detection and secret detection for LLM pipelines: block credentials, redact PII, and stop data loss before it leaves your perimeter. A practical.
Input vs. Output Guardrails: Where to Enforce What
A practical mental model for AI guardrails architecture: where input guardrails and output filtering belong, which checks run on both sides, and latency.
Designing for LLM Reliability: Failing Safe Under Load
Engineer production LLM reliability: fail-safe vs fail-closed by severity, a stateless hot path, explicit latency budgets, high availability AI guardrails.
Self-Hosted vs. Hosted Guardrails: Choosing a Deployment Model
Self-hosted AI guardrails vs. a hosted control plane: how the deployment model shapes data residency, latency, compliance, and operations for production LLM.
Streaming Guardrails: Enforcing Policy on Token Streams
Streaming guardrails for LLM apps: buffering strategies, real-time moderation, and output filtering that keep token streaming fast without leaking unsafe.
Jailbreak Detection: Patterns, Limits, and Layered Defense
A practical guide to LLM jailbreak detection: common adversarial patterns like DAN, why detection has limits, and how defense in depth with AI guardrails holds.
Measuring Guardrail Quality: Precision, Recall, and Latency Budgets
Learn guardrail evaluation with precision and recall, false positives, latency budgets, and coverage. A metrics-driven guide to benchmarking AI guardrails.
Building a Guardrail Policy: From Default-Deny to Production
Learn how to build a guardrail policy as code: start from severity tiers, use default-deny, monitor before blocking, scope per environment, and version with.
Observability for AI Guardrails: Traces, Metrics, and Audit
AI guardrails observability done right: LLM tracing, guardrail metrics, and tamper-evident audit logs with a metadata-first approach for trustworthy LLM.
Tool-Use Governance: Securing Function Calling and Agents
Tool governance for AI agents: allowlist tools, validate function calling arguments, add human-in-the-loop, and audit every call. Practical LLM tool use.