#ai-defense
-
Indirect Prompt Injection Explained: Defenses That Hold
How indirect prompt injection works, why LLMs cannot separate instructions from data, and the layered defenses Microsoft and Google use.
-
LLM Guardrail Benchmarks: Build Your Own Eval Set
Why published guardrail benchmarks do not transfer, and how to build a held-out eval set with hard negatives, attack success rate, and a latency budget.
-
LLM Audit Logging: What to Log, Redact, and Retain
A field-level guide to LLM audit logging: which attributes to persist, what to redact before storage, retention tiers, and the compliance floor.
-
Prompt Injection Detection Methods: A Practitioner's Guide
A comparative guide to prompt injection detection methods: classifier scanning, representation analysis, canary tokens, and output monitoring.
-
Monitoring LLM Outputs in Production: Anomalies and Drift
How to build production observability for LLM outputs, covering anomaly detection, latency alerting, output drift signals, and sane alert routing.
-
Output Filtering Architecture for Production LLMs
How to architect a multi-layer output filtering pipeline for production LLMs: deterministic guards, classifiers, schema validation, and async sequencing.
-
LLM Output Filter Latency: Budgeting Each Layer
What each output-filter layer costs in latency, from sub-millisecond regex checks to LLM-as-judge, and how to sequence them inside a real p95 budget.
-
Prompt Injection Prevention: Defense in Depth
A systems-level guide to preventing prompt injection in production LLMs, covering layered controls, structural prompt design, and privilege separation.
-
System Prompt Hardening Against Prompt Injection
How to harden an LLM system prompt against injection: explicit trust models, instruction hierarchy, and privilege separation between model roles.
-
AI Defense Techniques for LLMs: A Practitioner's Guide
A technical breakdown of LLM defense techniques, from input guardrails and prompt hardening to dual-model architectures, mapped to OWASP and NIST.