# AI Guardrails

> Understand what can go wrong. Put checks where they can stop it. Test what gets through.

Practical guides with source documents, worked examples and downloadable worksheets. Each guide includes its review date.

## Guides
- [What are AI guardrails?](https://aiguardrails.dev/text/what-are-ai-guardrails.md): Learn where AI guardrails check content, data and actions, and how to choose controls for a real application.
- [LLM guardrails for applications that generate answers](https://aiguardrails.dev/text/llm-guardrails.md): Design LLM guardrails around retrieval, generation, output validation and failure handling without assuming one filter covers every path.
- [Prompt injection prevention for apps and agents](https://aiguardrails.dev/text/prompt-injection-prevention.md): Reduce prompt injection risk by separating untrusted content from authority and testing the data and actions an attacker could reach.
- [AI agent guardrails for work with real consequences](https://aiguardrails.dev/text/ai-agent-guardrails.md): Control an agent's tools, targets and approvals so a mistaken plan does not automatically become an authorized action.
- [Tool calling guardrails that hold at execution time](https://aiguardrails.dev/text/tool-calling-guardrails.md): Validate tool arguments, check trusted permissions and bind approvals before executing an AI agent's proposed operation.
- [How to test AI guardrails](https://aiguardrails.dev/text/ai-guardrails-testing.md): Build a repeatable guardrail test suite that measures prohibited outcomes, valid-task failures and behavior during outages.
- [Guardrails vs fine-tuning: which problem are you solving?](https://aiguardrails.dev/text/guardrails-vs-fine-tuning.md): Choose between runtime checks and model training by separating task-quality problems from permissions and application rules.
- [How to choose AI guardrails tools](https://aiguardrails.dev/text/ai-guardrails-tools.md): Compare guardrail tool categories by what they inspect, where they enforce rules and how they fail in your application.
- [PII redaction in AI applications](https://aiguardrails.dev/text/pii-redaction.md): Place personal-data redaction before disclosure, test the entities your application handles and inspect tool and logging paths separately.
- [Structured output validation for AI applications](https://aiguardrails.dev/text/structured-output-validation.md): Validate generated JSON before using it, then check the business rules and permissions a schema cannot decide.
- [Fail-closed guardrails without accidental bypasses](https://aiguardrails.dev/text/fail-closed-guardrails.md): Define how protected actions stop when guardrail decisions are missing, stale or unusable, and plan recovery without granting wider permission.
- [An AI guardrails checklist for release review](https://aiguardrails.dev/text/ai-guardrails-checklist.md): Use an evidence-based checklist to review data access, outputs, tool permissions, failure handling and ownership before releasing an AI workflow.
- [Indirect prompt injection](https://aiguardrails.dev/text/indirect-prompt-injection.md): Trace an injected document or tool result through an agent, and test the boundary that prevents an unauthorized effect.
- [Prompt injection detection](https://aiguardrails.dev/text/prompt-injection-detection.md): Evaluate detector coverage, score thresholds, false blocks and outages against the inputs your application actually processes.
- [RAG security from ingestion to answers](https://aiguardrails.dev/text/rag-security.md): Preserve document permissions through chunks, retrieval, caches and deletion, with a worked two-tenant example.
- [AI agent sandboxes](https://aiguardrails.dev/text/ai-agent-sandbox.md): Review filesystem, network, identity and secret boundaries before giving a coding agent an execution environment.
- [LLM red teaming for agents](https://aiguardrails.dev/text/llm-red-teaming.md): Plan an authorized adversarial exercise, inspect real effects and turn findings into regression tests.
- [Prompt injection vs jailbreak](https://aiguardrails.dev/text/prompt-injection-vs-jailbreak.md): Distinguish instruction attacks, safety bypasses and tool effects so the test measures the failure you care about.
- [Memory poisoning in AI agents](https://aiguardrails.dev/text/memory-poisoning.md): Track untrusted information from memory write to later retrieval, correction and action.
- [LLM security: map the whole application](https://aiguardrails.dev/text/llm-security.md): Build a threat model around data, retrieval, tools, credentials and output consumers.
- [AI guardrails examples](https://aiguardrails.dev/text/ai-guardrails-examples.md): Work through schema, redaction, permissions, approvals and outage checks with explicit expected outcomes.

## Tools and resources
- [Try a policy check](https://aiguardrails.dev/tools/policy-checker/)
- [Worksheets and examples](https://aiguardrails.dev/resources/)

## About
- [Editorial policy](https://aiguardrails.dev/about/)