Resources
Read before you buy
The material around the product: what the problems actually are, the vocabulary to discuss them, how in/guard/out fits the stack you already run, and what regulators expect of LLM traffic. Written for engineers deciding what to defend against – not just what to buy.
Explainers
What are LLM guardrails?
LLM guardrails explained: input, output, and action checks that keep AI apps safe. Types, concrete examples, library vs proxy trade-offs, and how to start.
What is prompt injection?
Prompt injection explained: direct jailbreaks, indirect attacks hidden in emails and tool results, and why agents raise the stakes. Defense in depth, layer by layer.
What are AI hallucinations?
Why LLMs hallucinate, the four failure types that matter in production, and how detection actually works – from deterministic number checks to fact-check judges.
What is MCP security?
The Model Context Protocol gives agents tools – and attackers a surface. Tool poisoning, rug pulls, and result injection explained, plus the gateway defenses.
AI hallucination examples: the incidents that made case law
Documented AI hallucination incidents – Air Canada’s chatbot, fabricated legal citations, a $100B demo error – and the guardrail that would have caught each one.
What is excessive agency?
Excessive agency – OWASP LLM06 – is the gap between what your agent can do and what it should. Permissions, tiers, budgets, and approvals, explained.
The LLM security checklist
A practical pre-launch security checklist for LLM apps and agents: input, output, action, and operational checks – each with what “done” actually means.
The AI security challenges of 2026
The AI security problems defining 2026 – AI agents that act, MCP supply chains, hallucinated facts, and data leaks – and how runtime guardrails handle each one.
Do local LLMs have guardrails?
Local LLMs ship refusal training, not runtime guardrails. What Llama, Mistral, Gemma and DeepSeek actually enforce – and the guard models you can run yourself.
Agent inner loop security
An AI agent works in a loop: think, use a tool, read the result, think again. That loop is where things go wrong – and where the checks have to happen.
Inside-the-loop vs outside-the-loop
Should safety checks run inside the AI agent, or at a door it has to pass through? What each one sees, what each one misses, and how to split the job.
Model Context Protocol security advisories in 2026
A plain-English roundup of 2026 MCP security advisories – the NSA’s guidance, the year’s critical CVEs, the pattern behind them, and what to do about it.
Reference
The guardrails glossary
Thirty terms in a paragraph each – the attacks, the defenses, and the operating concepts – each linking to the page that treats it in depth.
OWASP LLM Top 10 coverage map
Each of the ten risks mapped to the guardrail that addresses it – including the two we only partially cover and the one a runtime proxy honestly can’t.
Compare
The four deployment shapes compared honestly – plus head-to-head pages against specific products: LiteLLM, Bifrost, Runlayer, MintMCP, the hallucination detectors, and more.
Integration guides
LangChain
LangChain 1.0 ships PII and human-in-the-loop middleware – wired per agent, in-process. in/guard/out enforces on the wire: every chain covered, PII restored, budgets across the run.
n8n
n8n’s Guardrails node checks text at nodes you place it. in/guard/out guards the whole workflow: every model call, checkpoints before consequential steps, your export compiled into policy.
MCP
The 2026 MCP spec added OAuth 2.1 authorization and tool annotations – it governs who may connect, not what calls may do. The gateway enforces the second half.
OpenAI
OpenAI ships a free Moderation API and an in-process guardrails library. in/guard/out runs the checks on the wire – PII restored, answers grounded, tool calls policed – via one base-URL swap.
Claude (Anthropic)
Point your Anthropic SDK at in/guard/out – native /v1/messages support. PII screening, fact-checking, and full agent tool policy on every Claude call, one base URL.
OpenRouter
OpenRouter’s built-in guardrails screen requests with regex and cap spend. in/guard/out adds the response side: multilingual injection defense, PII restore, fact-checking, agent enforcement.
LM Studio & local models
Local models still hallucinate and still obey injections. Run the same guardrail pipeline in front of LM Studio and any OpenAI-compatible local server.
Compliance mappings
EU AI Act
The 2026 digital omnibus moved the EU AI Act’s high-risk deadlines to Dec 2027 / Aug 2028 – but Article 50 transparency still lands 2 August 2026. What holds, and how runtime guardrails help.
GDPR
Sending prompts to model providers is processing personal data. How PII screening, pseudonymization, and screened logs support GDPR-aligned LLM deployments.
HIPAA
Using LLMs with patient data raises PHI disclosure risk. How PII screening and audit trails support HIPAA-aligned healthcare AI deployments.
Done reading? See it run.
We’re running a limited demo – sign up and we’ll get you in as soon as we can.