Key takeaways

Only 14 percent of organizations say they have formal GenAI policies that are actively enforced and audited, according to Check Point’s 2026 Cloud Security Report.

  • →The costliest AI mistakes break rules no vendor ships. Air Canada was held liable for a refund its chatbot promised and the airline never offered
  • →The usual fixes fall short. Models can be argued out of system prompts, keyword lists miss meaning, and US-centric DLP lists miss the data your regulators protect
  • →Custom guardrails turn a sentence into an enforced rule. Describe it, test it in under five minutes, and run it on every message and tool call in tens of milliseconds

That is why the rules only your business knows need the same enforcement as the threats everyone shares. See how Check Point AI Security protects AI apps and agents →

Custom guardrails, part of Check Point AI Security for securing AI applications and agents, let you write your own AI rules in plain language, test them and enforce them on every message and tool call. Every guardrail product ships with roughly the same list of threats, but the rules that get a business into trouble are its own: the advice its assistant must never give, or the data its regulators protect that standard data loss prevention (DLP) lists miss. This article shows how to write those rules yourself, DLP rules included.

The Rules Nobody Ships

Some of the most expensive chatbot mistakes of recent years involved no attacker at all. In 2024, a Canadian tribunal held Air Canada liable after its chatbot told a grieving passenger about a bereavement refund the airline did not offer, and it rejected the airline’s argument that the chatbot was responsible for its own words. That spring, New York City’s assistant for small businesses was caught telling employers they could take a share of their workers’ tips, and the city shut it down this year. And the parcel firm DPD found its chatbot swearing at a customer and calling DPD the worst delivery firm in the world. Each broke a rule its owner could have written in one sentence, and no vendor’s list contains it.

The same gap runs through DLP. Each country’s law sets its own categories of protected data, and they bind every business operating there, not only the government. US health data has 18 identifiers to remove before it counts as de-identified, Korea’s privacy law strictly limits the processing of resident registration numbers, and a bank in Spain has to catch DNI and NIE numbers, eight digits and a check letter, that no US list contains. Covering regulations like these has meant waiting for a vendor to add them, or reading them yourself, listing every data type they name and writing a rule for each one.

Why the Usual Fixes Fall Short

Teams usually reach for one of three fixes. The first is a line in the system prompt, such as never give investment advice. That is a request to the model rather than a control, since the model can be argued out of it and nothing records when it slips. The second is a list of keywords or regular expressions, which catches exact formats but not meaning, and misfires in both directions. Advice rarely contains the word advice, and one bank’s chatbot scolded a customer for typing “virgin” in a question about Virgin Money’s own savings accounts. The third is to ask a second AI model to judge every message. That understands meaning but adds hundreds of milliseconds and a cost to every message, and hands each one to a model provider. Teams need the understanding of the third approach at close to the speed of the second.

Describe It, Test It, Then Deploy It

Building a custom guardrail starts with a sentence. You describe what you want caught, and the builder’s AI assistant asks a question or two if the scope is unclear, then compiles a policy you can read before you save anything. It decides which kind of rule each part needs. Rules about meaning, such as whether a reply gives financial advice, catch paraphrases, work across languages and come with three sensitivity levels. Rules about exact formats, such as an IP address, an internal account number or a national ID, become pattern rules with nothing to tune. Logic joins the two: flag this, unless that is also present.

The rule in the figure needs all three: a pattern for the competitors’ names, a meaning rule for steering, and logic that flags a reply only when both appear. A reply that mentions a rival’s outage passes, and so does an upsell of your own plan.

A business rule written as one sentence, "Flag replies that steer customers toward a competitor," compiled into a pattern for the named competitors, a meaning rule for encouraging a switch, and logic that flags a reply only when both are true. Of three test replies, only the one that names a competitor and encourages switching is flagged.

Testing happens in the same conversation: you paste in example messages from your own app, or let the builder generate some, and see whether each was flagged and which rule fired. When a result is wrong, you say what should change and the policy updates. Getting from a description to a tested detection takes under five minutes, and one saved guardrail can serve many policies.

Fast Enough to Check Everything

The AI model only helps you build the guardrail. Once saved, the guardrail runs without a large language model in the loop, so the check is small, fast and makes no external calls. A policy with one or two rules adds about 35 to 40 milliseconds at the 95th percentile, so it can run on every message rather than a sample. From November 2026, it can also run in self-hosted environments, so the check stays inside your own infrastructure.

Rules for Agents, Not Just Chat

A custom guardrail is not limited to chat. Each one can be scoped to user messages, model replies, or the inputs and outputs of an agent’s tool calls, so one rule can follow data through an agent’s whole workflow. A DLP rule for national ID numbers can check what a support agent writes into a ticket or an outbound email, not only what it says on screen. A pattern that would be noisy in conversation, such as a command to drop a database table, becomes a precise check once it applies only to a database tool. Last week’s piece on Dangerous Deviation Detection asked whether an agent’s action is dangerous and unwarranted. A custom guardrail asks a narrower question at the same boundary.

Does this break a rule only we would know to write?

Where It Fits and Where It Does Not

Custom guardrails work best on one bounded topic at a time, whether a kind of advice, a data format or a set of named companies. They check one message at a time, so they will not catch something that only becomes a problem across a conversation. They are also not meant for adversarial threats, which the prompt attack detectors in the same policy are built for. Some requirements stay genuinely hard, such as judging whether generated code is safe. And because some patterns overlap with ordinary language, the best first step is to run a new guardrail in detect mode and watch what it flags before it blocks anything.

Your Rules, Checked Like Everyone Else’s

Every business already knows the rules its AI has to keep, and until now enforcing them meant a vendor’s roadmap or a page of regular expressions. Custom guardrails turn each rule into a sentence you can test in minutes and enforce on every message and tool call, so the rules only your business knows get the same protection as the threats everyone shares.

This piece is part of a series on protecting AI agents beyond the classic guardrail. Next week, we look at audio guardrails and at what changes when an agent’s input is a voice.

Get started

Enforce the Rules Only Your Business Knows

No vendor’s list of threats can cover every rule a business lives by. Custom guardrails let you describe your own in plain language, test them on your own examples, and enforce them on every message and tool call in milliseconds. They are part of Check Point AI Security, which shows you the agents running across your business, assesses each one’s risk in context, and controls what they can access and do at runtime.

Demo

Book a session with our team to build a guardrail for your own use case.

Book a session →

Docs

Build your first custom guardrail and test it on your own examples.

Read the docs →

Webinar

Watch our session on securing AI apps and agents.

Watch the session →

You may also like