AI data loss prevention
Keep secrets and PII out of AI traffic
Ferentin scans prompts, tool calls and results for personal data, card numbers, health identifiers and secrets. It redacts, blocks or logs each match at the gateway and on the developer’s Mac.
The problem
Agents move data where no one reviews it
A coding agent reads a .env file while debugging and the key lands in a model provider’s context. An MCP tool returns a customer record nobody asked to see. DLP built for email, uploads and USB drives watches the exits people use, not the tool arguments and results an agent writes on its own.
How it works
One policy, enforced in two places
- 1
Choose what counts as sensitive
Pick detector profiles such as US PII, EU PII, Financial, Secrets and Healthcare, then untick any detector that misfires in your context. Validators such as checksums confirm a match before it counts. Paste a prompt or a tool result into the tester to see what each detector catches before you save.
- 2
Decide what happens to each match
Set Redact, Block or Log only per category, with exceptions for single detectors. At the gateway, Redact swaps a value such as an email address for a typed placeholder like [EMAIL_1] and restores it in the model’s response. SSNs and card numbers are masked to their last four digits. Block stops the request and returns your message.
- 3
Enforce at the gateway and on the Mac
At the gateway a policy scans LLM input, model output and MCP tool results. On enrolled Macs it checks tool calls before they run in Claude Code, Cursor, Codex and GitHub Copilot; prompts in Claude Code, Cursor and Codex; tool output in Claude Code and Codex. The device scan runs locally, so matched text never leaves the Mac.
- 4
Read what was caught
Every gateway event records which detectors matched, the effect applied and the match count. If you keep payloads for audit, the stored copy is sanitized too, so a reviewer sees XXX-XX-1120 instead of the full number. Activity lists what detectors caught on managed Macs and whether it was stopped.
Works with your stack
Works with your stack
One policy covers the models and MCP servers behind the Ferentin gateway. The same policy reaches the coding agents on your Macs.
DLP for AI detects, redacts or blocks sensitive data such as card numbers, government identifiers and credentials as it flows into and out of AI systems. It covers prompts and model responses, the arguments an agent passes to tools and the results that come back. Classic DLP watches email, uploads and USB drives. AI DLP has to sit where agent traffic flows, because agents move data on their own initiative.
Go deeper
Learn more
- Data loss prevention for AIGlossary: why DLP for AI has to scan tool arguments and results, not only prompts.
- How AI agents changed data protectionAgents read metadata and decide on their own. That moves the threat upstream.
- The Mac is the new AI perimeterHow Ferentin keeps secrets out of prompts, tool calls and tool output on the Mac.
Related
- AI & MCP GatewayOne governed path to every model and MCP server, managed or in your cloud.
- EndpointsA managed macOS agent that finds and governs the AI tools on every Mac.
- Agent Runtime SecurityEvaluate each agent action in context. Stop toxic flows before they run.
- AI ComplianceSigned evidence of every AI call, policy decision and approval.
See it on your own stack
Walk through it with our team on your identity provider, your models and your MCP servers. Or start free today.