Models suggest. Your system decides.
AI guardrails,
from prompt
to production.
Understand prompt injection, validate outputs and control tool actions. Follow worked examples, then test what your own system allows.
- 01Input & retrieval
A request meets outside content
Keep the user’s authority separate from documents and tool results.
Trace an indirect injection ↗ - 02Model output
The model proposes an action
Check the structure and meaning of the proposed arguments.
Validate the output ↗ - 03Application boundary
Enforce identity, scope and approval
The system receiving the action checks whether it is allowed.
Control the tool call ↗ - 04Observed result
Check what actually happened
Test the protected resource and keep the decision record.
Build a useful test ↗
What are AI guardrails?
AI guardrails are checks on an application’s inputs, outputs and actions. A content filter can reject a response; an authorization check can reject a tool request. Each control protects a different boundary, and each needs its own failure tests.
Choose your next task
Understand it. Build it. Check it.
Foundations
Understand the layers
Separate model behavior, application checks and authority over real systems.
All 3 foundations guides →Threats
Follow the untrusted input
Trace an instruction through documents, tool results and persistent memory.
- Indirect prompt injection
- Prompt injection prevention for apps and agents
- Memory poisoning in AI agents
Implementation
Put the check where it holds
Validate a response, scope a tool call and limit an agent’s environment.
- Tool calling guardrails that hold at execution time
- RAG security from ingestion to answers
- AI agent sandboxes
Testing & tools
Test the actual outcome
Use failure cases and repeatable evidence to evaluate the controls you chose.
All 6 testing & tools guides →Work through a concrete decision
The example is where the rule gets specific.
Follow a request through the control, the failure case and the evidence you would inspect. The examples make their assumptions explicit.
Open the worked examples ↗Use what you learn
Try a policy decision.
Change the request, access and approval state. See why an illustrative policy allows, holds or denies the action.