ailp

Inspiration

AI adoption is moving faster than enterprise security controls. Employees want to use AI assistants for productivity, but security teams need to prevent confidential information from being shared with external AI services.

Traditional DLP solutions were built mainly for email, web uploads, and endpoint files. AI prompts create a different challenge: the content is unstructured, multilingual, conversational, and often contains sensitive data mixed with normal business context.

We built AILP to give organizations a practical control point between their users and AI services—so they can adopt AI without treating data protection as an afterthought.

What it does

AILP is an ICAP-based AI Data Loss Prevention gateway. It inspects text and supported files before they are sent to an AI service.

It detects sensitive data such as personally identifiable information, credentials and secrets, financial identifiers, internal URLs, certificate material, and other confidential content. Based on policy, AILP can log the event, block the request, or redact sensitive values before the request continues.

The platform also records policy decisions and operational data so security teams can understand what is being detected, which policies are triggered, and where protection needs tuning.

How we built it

We built AILP as an inline ICAP gateway that can integrate with existing enterprise traffic-control infrastructure. The detection layer combines deterministic rules for structured identifiers, such as IBANs, phone numbers, and email addresses, with locally deployable privacy models for language-dependent content such as names, addresses, dates, and secrets.

The system supports text inspection and file-content analysis. Its policy engine evaluates findings and applies the configured action: log, block, or redact. The architecture is designed to support local deployment, reducing the need to send sensitive content to another external inspection service.

We also designed it with operational constraints in mind, including response timeouts, fallback behavior, event logging, and dashboards for security teams.

Challenges we ran into

The main challenge is balancing protection and usability. If detection is too broad, it creates false positives and interrupts legitimate work. If it is too permissive, sensitive data may leave the organization.

This is especially difficult for Turkish and mixed Turkish-English enterprise content. Names, addresses, dates, passwords, technical documentation, logs, and file-extracted text can appear in many formats. We also needed to handle imperfect document extraction and ensure the inspection path remains fast enough for real AI usage.

Accomplishments that we're proud of

We are proud that AILP moves beyond a proof of concept and addresses the actual deployment needs of enterprise AI security.

It provides a single enforcement point for prompt and file inspection, supports multiple policy actions, and combines fast structured detection with local model-based detection. It is also designed to work with existing ICAP-based enterprise environments instead of requiring organizations to redesign their AI traffic flow.

Most importantly, AILP gives security teams a way to enable AI use with visibility and control, rather than choosing between unrestricted access and a complete ban.

What we learned

We learned that AI DLP is not only a detection problem. Accuracy, latency, fallback behavior, explainability, and policy tuning all matter in production.

We also learned that language-specific evaluation is essential. A model that performs well on English benchmark data may still produce unacceptable false positives or missed detections in Turkish enterprise content. Realistic test data, production logging, and continuous feedback are necessary to improve the system safely.

What's next for AILP

Next, we will expand production-style evaluation using logged policy outcomes and analyst feedback. We plan to improve detection accuracy for Turkish and mixed Turkish-English content, broaden file-format coverage, and make policy reporting more useful for security teams.

We also want to make deployment simpler, improve performance for CPU-based environments, and provide clearer controls for organizations that need to apply different policies by user group, data type, or AI service.

Built With

• fastapi
• go
• pydantic
• python
• sqlalchemy