All articles
Reviews of AI security products and platforms built from the public record: vendor documentation, published evaluations, source code where the tool is open, and independent results where they exist. Every claim is attributed and the limits of the evidence are stated.
-
How to Secure Vector Database Access in RAG Systems
Vector stores in RAG pipelines carry auth gaps, embedding inversion risk and cross-tenant exposure. How to lock down both the read and write path.
-
AI Security Review: A Checklist for LLM Systems
An AI security review checklist for LLM systems: trust boundaries, data flows, control mapping, evidence to demand, and the findings reviewers miss.
-
LLM Security Benchmarks Compared: What Each Measures
LLM security benchmarks compared: what HarmBench, JailbreakBench, AgentDojo, AgentHarm and CyberSecEval each measure, and where every one of them stops.
-
RAG Pipeline Security Best Practices: A Checklist
RAG pipeline security best practices: retrieval poisoning, access control at query time, embedding risks, and how OWASP and MITRE ATLAS frame the threats.
-
LLM Guardrail Frameworks Compared: Six Options
A comparison of NeMo Guardrails, Guardrails AI, Bedrock, Azure Prompt Shields, Model Armor, and Lakera by OWASP coverage and deployment trade-offs.
-
How Prompt Injection Detection Works, Tier by Tier
How prompt injection detection works, from input scanning and embedding classifiers to fine-tuned guardrails and runtime monitoring, with the trade-offs.
-
Best AI Security Testing Tools 2026: Scanners and Red Teams
A practitioner's comparison of the best AI security testing tools in 2026: open-source scanners, commercial red-teaming platforms, and how to choose.
-
OWASP LLM Top 10 2026: Mitigations for Every Risk
A practical OWASP LLM Top 10 2026 mitigation guide covering all ten current risk categories, controls, implementation gaps, and layered defenses.
-
Patronus AI Review: Automated LLM Evaluation and Guardrails
A review of Patronus AI's evaluation platform: the Lynx hallucination model, the Glider custom evaluator, built-in safety checks, and published pricing.
-
Protect AI's ModelScan and NB Defense: An Open-Source Review
A review of Protect AI's two best-known open-source tools — ModelScan for model serialization attacks and NB Defense for Jupyter notebooks.
-
Robust Intelligence (Now Cisco AI Defense): Platform Review
A review of Robust Intelligence, now part of Cisco AI Defense: algorithmic red teaming, model file scanning, and runtime protection of AI applications.
-
Giskard Review: Open-Source Testing for LLM and RAG Apps
A review of Giskard, the open-source Python library for testing AI systems, covering its automated LLM vulnerability Scan and the RAGET RAG toolkit.
-
How to Evaluate AI Security Tools Without Getting Fooled
AI security tool demos are built for best-case scenarios. The evaluation dimensions, protocol, and vendor questions that expose how a tool really performs.
-
PyRIT Review: Microsoft's AI Red Teaming Framework
A review of PyRIT 1.1, Microsoft's open-source AI red teaming framework: its target, attack, converter, scorer and memory design, and multi-turn attacks.
-
Guardrails AI: Output Validation Without Retraining
Guardrails AI provides a validation layer for LLM outputs — checking format, structure, and content without touching the model.
-
Arize Phoenix: LLM Observability That's Actually Free
Arize Phoenix is a source-available (ELv2) LLM observability platform that grew beyond its drift-detection origins, with evaluation tooling built in.
-
Garak LLM Scanner Review: Research Tool or CI Gate?
A review of garak, NVIDIA's open-source LLM vulnerability scanner: plugin architecture, backend coverage, report quality, and the CI-gating pattern.
-
Rebuff: Open-Source Prompt Injection Defense, Layer by Layer
Rebuff is a self-hosted prompt injection detector with four layers: heuristics, LLM-based detection, a vector database of past attacks, and canary tokens.
-
Lakera Guard: Prompt Injection Detection in Practice
Lakera Guard is built for prompt injection detection rather than content moderation. What the documentation shows about coverage, latency, and cost.