OverviewWhat it is
Guardrails are the controls around the model: blocking unsafe content, catching hallucinations, protecting private data, defending against prompt injection, and meeting policy and regulation. In most products, trust is the product - one bad output can end adoption.
At a glanceGuardrails & Responsible AI
Guardrails wrap the model on both sides. Design for the failure modes, not just the happy path.
CompareThe guardrail landscape, side by side
Every guard worth layering, grouped by where it acts — on the input, inside the model, or on the output. Filter by stage or search across threat, mechanism, and tooling. Real systems stack several; defense in depth beats any single check.
Every row has a page — what it does, what it costs you, and how to tell when it is the thing biting you.
MechanicsHow it works
Defence in depth: filter inputs (block malicious prompts, redact PII), constrain the model (system rules, grounding), validate outputs (safety, factuality, format), and route high-stakes cases to a human.
Ground levelWhat you actually build
Layers before, around and after the model — and the latency, cost and false positives each one buys you.
LandscapeTypes & approaches
Click a highlighted type to open its own page — concept, use case, and diagram.
FeasibilityArchitecture & feasibility
Architecture & feasibility
- Guardrails are layered subsystems (input filter, output validator, escalation) - a feasibility requirement for regulated or high-stakes domains.
- Prompt injection is the top risk once an agent has tool access; least-privilege tools and output validation are the mitigations.
- Each guardrail adds latency and cost; the architecture task is placing the minimum set that covers your real failure modes.
In practiceWhat it means for building
Responsible-AI thinking is table stakes, not a nice-to-have. Map the worst-case outputs for your use case and require controls before launch.
Layer input filters, output validators, and escalation. Design explicitly for failure modes and adversarial inputs, not just the happy path.
GlossaryKey terms
CheckCheck your understanding
How do you stop hallucinations?
No single fix: ground with RAG, constrain via prompts, validate outputs against sources, show citations, and keep humans in the loop for high-stakes cases.
What is prompt injection and why care?
Hostile text that overrides your instructions. It matters most for agents with tool access - mitigate with input filtering, least-privilege tools, and output checks.
How much guardrailing is enough?
Match it to the worst-case harm of your use case; place the minimum layers that cover those failure modes.
What changedWhat changed here
As of 2026-09-25 — injection and guardrail coverage in the daily brief: 3 items in the last 7 days
All 3, newest first, from the Signal daily brief
Updated this page Microsoft published a humanist AI code of conduct and opened a six-week public consultation on the draft.
Updated this page Claude outputs will now carry detectable watermarks, affecting provenance and downstream redistribution.
Updated this page Claude will watermark AI-generated text, so pipelines should expect provenance checks and shifting user expectations around disclosure.
Updated this page Claude output is now invisibly watermarked worldwide; builders on Claude must plan for content-authenticity tooling.
Add watermarking and provenance metadata to the guardrails page as a compliance/transparency control for model output.
Updated this page EU AI Act transparency rules took effect in August, requiring disclosure for chatbots and labelling of AI-generated content.
Add to guardrails a section on the EU AI Act transparency obligations in effect from August 2, covering AI-interaction disclosure and synthetic-media labelling.
Three kinds of claim, strongest first. Signal runs every morning.