The Mavs Reins: AI Guardrails for Data Security That Enable Not Restrict

AI guardrails for data security usually mean a mechanism to allow, block or mask a prompt. This follows from the old style of data protection where DLP either allowed or blocked a request.

But a prompt is not a file moving from point A to point B, it is an input going into a model. To evaluate a prompt for data security, you need to understand its meaning, be precise about what is sensitive in it, and maintain continuity, because prompts are part of a conversation, not isolated events. The old DLP style cannot do this. It was built to find sensitive patterns in files moving between systems, not to read meaning across a conversation.

And even where evaluation lands, the actions themselves are damaging. Blocking kills the productivity of employees and agents, and masking reduces the quality of the model's response.

As work is shifting from databases and excel files to natural language used for interacting with AI, guardrails built for runtime AI security need to shift too, from blocking or masking to something smarter and more contextual, one that understands the meaning of a prompt, protects its context, and respects enterprise data sovereignty. This is the problem that the Mavs AI guardrails for data security solve.

Reins Do Not Constrain the Horse, They Give It Direction

The Mavs AI guardrails are built to enable productivity, not constrict it.

Just like the reins of the horse are a control mechanism to give it direction, not to restrain it, Mavs AI guardrails are the control mechanism that enterprises can use to secure prompts at runtime without constraining them.

Instead of simply blocking the request, the guardrails choose from a variety of policy actions according to the data in the prompt, keeping it safe to process while preserving its usefulness.

How Mavs Identifies Sensitive Data in AI Prompts

Every Mavs guardrail action begins by reading the meaning of a prompt as it evolves across the conversation, and uses that meaning to identify the sensitive data in it. The same string can be sensitive in one prompt and neutral in another, and the identification treats it accordingly.

What counts as sensitive goes beyond PII, PCI and PHI to include sensitive business information and Company Identifiable Information (CII): confidential project and deal codenames, unannounced pricing, M&A terms, unreleased roadmaps, key accounts, intellectual property, and trade secrets.

The Mavs Reins: AI Guardrails for Data Security that Enable Productivity

1. Smart Reprompt: Context-Preserving Data Protection for AI Prompts

When a prompt contains sensitive data, Mavs Smart Reprompt is the action of choice, a unique guardrail built natively for AI.

The idea behind Smart Reprompt is to protect the sensitive data in a prompt without taking away the context the model needs to generate an accurate response.

Once the sensitive information in a prompt has been identified, Mavs generates synthetic stand-ins that are similar in meaning and context. For example, the replacement can preserve attributes such as ethnicity, company domain, product category and other characteristics that matter to the model.

The prompt is then intelligently rewritten using these stand-ins, while keeping its original meaning, context and intent intact. This is one of the hardest parts of the process.

Finally, when the model responds, Mavs identifies the synthetic values in the response and swaps them back with the original values before returning it to the user.

The model sees the same context and meaning, but never sees the actual sensitive data. The user gets a response as if the original data had been shared with the model.

This matters because LLMs perform better when they have the complete context of a problem. Masking or removing information can degrade the quality of the response.

2. The Human Judge: User-in-the-Loop AI Guardrail

The Human Judge is the right action when the enterprise wants the user to decide, rather than acting automatically. It produces a notification when a user tries to send a prompt with sensitive data, acting as a moment of caution. The user can see what was flagged in their prompt as sensitive, understand how that violates AI policy, and still choose to override the notice and process the prompt as is, or make changes to their prompt.

This is a useful action for employees you trust to make the call and for use cases that might not always be sensitive but would benefit from some layer of protection. For governance, each instance of the notification and the override is recorded and auditable.

3. Sovereign Reroute: AI Data Sovereignty With Internally Hosted Models

Rerouting prompts to an internally hosted safe model is useful not just from a security and data sovereignty perspective, but also when enterprises have regulatory requirements, sensitive intellectual property, or contractual obligations around how data can be processed. Sovereign Reroute allows companies to keep highly sensitive information within their own environment while still using AI, rather than having to send that information to an external frontier model.

This is a powerful option because Mavs can automatically detect when Smart Reprompt is not sufficient to protect the sensitive data, and reroute the prompt to an internally hosted model. This can happen when the data is too intertwined with the prompt to substitute cleanly, or when enterprise or regulatory policy forbids the data from leaving the environment even in transformed form, as with certain regulated health or financial records.

This allows the task to continue safely without interrupting the user's productivity or impacting the application's performance.

4. The Watchman: AI Usage Audit and Governance

This action doesn't change anything in the prompt, but records each prompt with a full audit log.

This gives security and IT teams visibility into how AI is actually being used across the enterprise: what users are asking, where risks are emerging, and which AI tools are being used. These insights power Mavs AI's Risk and Productivity Dashboards, giving enterprises the AI governance visibility they need to meet compliance requirements, fine-tune AI policies, and measure AI adoption and productivity over time.

5. Ultimate Block: Defense Against Harmful Prompts and Prompt Injection

The preceding actions ensure that most prompts sent to GenAI produce a meaningful response while keeping the data protected.

Ultimate Block completes the control layer for cases where processing the prompt would itself be a violation, such as a prompt injection or an otherwise harmful prompt that shouldn't reach the model at all.

Together with the other actions, Ultimate Block ensures that malicious prompts are blocked while leaving room for productivity everywhere else.

The Reins in Practice

Together, Smart Reprompt, The Human Judge, Sovereign Reroute, The Watchman, and Ultimate Block form a layered control system that adapts to what is actually in a prompt. Each is a different way of directing the prompt rather than restricting it.

This is what AI-native guardrails look like: they read each prompt, choose the right action for its content, and keep both the data protected and the work moving. The old DLP style could not do that. Mavs was built to.

Mavs AI runtime guardrails give enterprises data protection, data sovereignty, and AI governance for prompts sent to any AI model. To see the guardrails handling your own prompts, book a demo.

Frequently Asked Questions

What are AI guardrails?

AI guardrails are runtime controls that inspect prompts before they reach an AI model and choose the right action based on what the prompt contains. Unlike traditional data loss prevention systems built for files, AI guardrails are built for natural-language prompts and the conversations they sit inside.

How is Mavs AI different from traditional DLP and AI DLP?

Traditional DLP was built to spot sensitive patterns in files moving between systems, using allow, block, or mask actions. AI DLP typically extends this pattern-based approach to AI prompts. Mavs AI guardrails address the same problem — protecting sensitive data in AI prompts — but with a fundamentally different approach: they are built natively for AI. Mavs reads each prompt in context, identifies sensitive data with precision, and chooses from a range of actions including context-preserving transformation, so prompts stay useful and models produce accurate responses.

What business-sensitive data does Mavs protect?

Business-sensitive data is any information whose exposure could harm the enterprise: intellectual property, trade secrets, unannounced pricing, M&A terms, unreleased roadmaps, confidential project and deal codenames, and key accounts. Mavs identifies it in prompts alongside PII, PCI, and PHI. You may also see this framed as Company Identifiable Information (CII), the business analog of personally identifiable information. More in Introducing Business Sensitive Data Detection.

What is prompt injection and how does Mavs prevent it?

Prompt injection is an attack that uses the input itself to make an AI model follow the attacker's instructions instead of the ones it was given. Mavs prevents prompt injection through the Ultimate Block action, which stops harmful and malicious prompts before they reach the model.

How do Mavs AI guardrails support data sovereignty?

Through the Sovereign Reroute action, Mavs can automatically detect prompts containing data that should not leave the enterprise environment and reroute them to an internally hosted AI model. This keeps highly sensitive data within the enterprise while still allowing employees and applications to use AI.

How does Mavs support AI governance?

Every prompt is recorded and auditable, and the resulting Risk and Productivity Dashboards give security and IT teams visibility into how AI is used across the enterprise. This is the foundation for enterprise AI governance: meeting compliance requirements, fine-tuning AI policies, and measuring AI adoption over time.

Where can enterprises deploy Mavs AI guardrails?

Mavs AI guardrails can be enforced wherever employees or applications interact with AI. Available surfaces include Mavs Secure Chat, a model-agnostic chat workspace that covers OpenAI, Claude, Gemini, LLaMA and other models; Claude Desktop secured by Mavs, a runtime gateway integration for teams already using Claude; the Mavs API, for applications and agents that need to route through the same controls; and a browser extension for AI in the browser.