Applying DLP to Prompts for ChatGPT and Claude

Applying Data Loss Prevention (DLP) to prompts is essential for preventing sensitive data from being sent to third-party LLMs. An AI gateway like Bifrost can enforce DLP policies, scan for secrets, and redact PII before prompts leave the corporate network, securing the use of tools like ChatGPT and Claude.

A primary challenge for enterprises adopting generative AI is preventing the accidental exposure of sensitive data. When employees use models like OpenAI's ChatGPT or Anthropic's Claude, any proprietary code, customer information, or intellectual property included in a prompt is sent to a third-party service. This creates a significant risk of data leakage, violating compliance standards like SOC 2, GDPR, or HIPAA. To address this, organizations implement Data Loss Prevention (DLP) strategies for AI, and a central point of enforcement is an AI gateway. Bifrost, an open-source AI gateway from Maxim AI, provides tools to inspect and control the data within prompts before they reach any model.

The Challenge of Sensitive Data in LLM Prompts

Large Language Models have no inherent understanding of what constitutes sensitive information. They process prompts as-is, which can lead to several data security issues:

• Training on Prompts: Some service providers may use submitted data to train future versions of their models. While major providers like OpenAI and Anthropic now offer enterprise agreements with zero-data-retention policies, this is not always the default and may not apply to all service tiers.
• Accidental Exposure: An employee seeking help debugging a function might paste a code snippet containing proprietary algorithms or embedded API keys. Similarly, a support agent might paste a customer conversation that includes Personally Identifiable Information (PII) to summarize it.
• Compliance Violations: The transfer of regulated data (such as financial information or protected health information) to a third-party AI service without proper controls can violate data protection regulations, leading to significant fines and legal consequences.
• Shadow AI: Employees often use personal or unmanaged accounts for AI tools on their work machines. This "shadow AI" usage bypasses any existing security controls, creating a major blind spot for IT and security teams.

Methods for Prompt DLP Enforcement

Organizations use several methods to apply DLP to AI prompts, each with different points of enforcement and effectiveness.

1. User Training and Manual Policies

The most basic approach is to establish clear internal policies about what data can and cannot be used in prompts and to train employees on these guidelines.

• Pros: Low technical overhead to implement.
• Cons: Relies entirely on human vigilance and is prone to error. It is not a preventative control and offers no technical enforcement or audit trail.

2. Browser Extensions and Endpoint Agents

Some solutions use browser extensions or simple endpoint agents to scan text entered into web-based AI interfaces.

• Pros: Can catch data before it leaves the browser.
• Cons: Limited to specific web pages (e.g., chat.openai.com) and easily bypassed. They do not cover API-based integrations, desktop applications like Claude Desktop, or CLI-based coding agents.

3. Integrated DLP in AI Platforms

Certain enterprise AI platforms include built-in DLP features. These are often tied to the specific platform and may not cover the full spectrum of AI tools an organization uses.

• Pros: Tightly integrated with the platform's workflow.
• Cons: Creates vendor lock-in and does not address the use of other, unmanaged AI services.

4. Centralized AI Gateway

An AI gateway acts as a single point of entry for all LLM traffic. By routing all requests through a central proxy, security teams can inspect every prompt from any source (API, desktop app, custom integration) and apply consistent DLP policies.

• Pros: Provides universal, consistent policy enforcement across all AI applications and models. Offers detailed audit logs for compliance.
• Cons: Requires routing all AI traffic through the gateway, which needs to be highly performant to avoid adding latency.

How an AI Gateway Enforces Prompt DLP

An AI gateway like Bifrost is a powerful control point for implementing robust DLP. It inspects every request in transit and applies a series of security checks, known as guardrails, before forwarding the request to the upstream LLM provider.

The process is transparent to the end-user. A developer making an API call or an analyst using ChatGPT Desktop interacts with their tool as usual. The gateway intercepts the traffic and enforces policy in real-time.

Key DLP Features in Bifrost

Bifrost provides specific guardrail capabilities designed to detect and handle sensitive data within prompts.

• Secrets Detection: The gateway includes a built-in secrets detection guardrail that scans prompts for patterns matching API keys, database credentials, private keys, and other tokens. If a secret is found, the request can be blocked entirely, and an alert can be generated. This prevents credentials from accidentally being exposed to a third-party model or logged in its systems.

• PII Redaction with Custom Regex: Organizations can define their own patterns for sensitive data using the custom regex guardrail. This is commonly used to find and redact PII like social security numbers, credit card numbers, phone numbers, or internal project codenames. The gateway can either block the request or redact the sensitive portion, replacing it with a placeholder before it is sent to the LLM.

• Third-Party Guardrail Integration: Beyond its native capabilities, Bifrost integrates with specialized AI security providers like AWS Bedrock Guardrails, Azure Content Safety, and Patronus AI. This allows teams to apply sophisticated, pre-built content policies for PII, toxicity, and other sensitive topics directly at the gateway layer.

• Comprehensive Audit Logs: Every request and its outcome are recorded in immutable audit logs. This provides a complete trail for security reviews and compliance reporting, showing which prompts were sent, which were blocked by DLP policies, and why.

Extending DLP to the Endpoint with Bifrost Edge

A gateway can only enforce policies on traffic it sees. The rise of desktop AI apps and CLI agents creates a "last mile" problem where employee machines generate AI traffic that may not be routed through the central gateway.

To solve this, Bifrost's governance and security controls are extended to the device level with Bifrost Edge. Edge is an endpoint agent that runs on employee machines (macOS, Windows, and Linux) and automatically routes all AI traffic from supported desktop apps, browsers, and coding agents through the organization's Bifrost instance. This ensures that the same DLP guardrails and security policies apply to prompts from tools like ChatGPT Desktop or Claude Desktop, closing the shadow AI loophole.

Implementing a Secure AI Workflow

For organizations looking to leverage the power of tools like ChatGPT and Claude securely, applying DLP at the prompt level is not optional. While user education is a necessary first step, technical enforcement is required for reliable protection. An AI gateway offers a centralized, scalable, and comprehensive solution for inspecting and controlling data before it leaves the network perimeter.

Teams evaluating AI security solutions can request a demo of Bifrost to see how its guardrails and endpoint governance can be applied or review the project's open-source repository.