Guardrails
Guardrails are the checks and limits wrapped around an AI model so its behaviour stays inside your rules: they screen what goes in, constrain what the model can do, and verify what comes out before it reaches a person or a system.
No jargon, no math
In plain words, what is it?
A bare model will answer almost anything fluently; guardrails are the fence around it. They work in layers. Before generation, input filters scan the prompt for injections, leaked personal data or requests outside the system's scope. During generation, the system prompt, the allowed tools and parameter limits bound what the model may touch. After generation, output checks score the answer against your policies and its sources: does it contain personal data, unsupported claims, off-policy promises? Anything that fails gets blocked, rewritten or handed to a human, and the event is logged. The mechanism is the same as airport security: every bag passes the same checks, whoever owns it. Where it breaks: rules tuned too tight bury users in false blocks, too loose let bad content through, and a genuinely novel attack pattern can slip past pattern-matching filters. That is why high-stakes actions get a human gate anyway.
- 01
Screen the input. Prompts are scanned for injections, personal data and out-of-scope requests.
- 02
Constrain the model. System prompt, allowed tools and limits bound what it can do.
- 03
Check the output. The answer is scored against policy and its sources.
- 04
Block or rewrite. Failing content never reaches a user unchanged.
- 05
Log and escalate. Every block is recorded; repeated or risky attempts go to a human.
Terms that live nearby.
Prompt Injection
Prompt injection is an attack where hidden instructions are planted inside content an AI reads (an email, a web page, a document), so the model executes a stranger's commands as if they were yours.
ReadJailbreak
A jailbreak is an attempt to make an AI model ignore its safety training through a cleverly disguised prompt, so it produces content it was built and tuned to refuse.
ReadSystem Prompt
The system prompt is the standing instruction a language model receives before any user message: it defines the model's role, tone, limits, output format and the tools it may use, and it shapes every answer in the conversation.
Read§ 04
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.