Agents read everything they're allowed to
A copilot inherits a user's permissions. If a folder is overshared, so is every answer built from it. Nobody reviews each read.
DLP for AI agents & copilots
AI agents move your data. Haltgate decides what leaves.
Real-time control over the sensitive data that Microsoft Copilot, ChatGPT and Claude connectors, MCP servers and your in-house agents read and send out. Haltgate finds PII and secrets in flight, then allows, warns, redacts or blocks them, and keeps the evidence your auditors ask for.
01 The problem
Your data now has a new way out. It doesn't ask first.
People used to be the leak. Now it's the software acting for them. Copilots summarize SharePoint, connectors pull CRM records into chats, and agents call tools that post data to the outside world. Most of it happens where legacy DLP can't see.
Data leaves in tool calls, not pastes
MCP servers and agent tools move records as JSON, through APIs, webhooks and outbound requests. There's no paste event for an extension to catch.
Legacy DLP was built for email and USB
Hand-written rules and keyword lists drown teams in alerts and don't know which agent acted for which user. Auditors still want proof.
35%
of AI-using small businesses found unapproved AI tools, workflows or agents running in the past six months.
Source: Pax8 survey of 1,150 SMBs, Oct 2026 ↗~60%
of the AI risks one security vendor found sat in tools that non-developers built themselves, with finance the biggest source.
Source: Pluto Security via TechCrunch, Oct 2026 ↗$2B+
spent by eight major security vendors buying "security for AI" startups in about 18 months. The industry knows this gap is real.
Source: Forrester ↗02 Try it
Type like an agent would. Watch the gate.
Paste a prompt, a tool call or a config file. Detection runs entirely in your browser, and nothing you type leaves this page.
output what the model receives
Demo detection uses pattern matching with checksums (Luhn for cards, mod-97 for IBANs). The product adds context-aware classification, identity and data lineage. All sample values are public test data.
03 How it works
Four moves, every time data tries to leave.
-
01
Detect
Inspect prompts, file reads, tool calls and responses in flight. Find PII, PHI, card data, credentials and API keys, including inside JSON and code.
email · iban · mrn · aws_key · jwt -
02
Decide
Weigh what the data is with who it belongs to, which agent is acting for which user, and where it's headed. You set policy in plain terms, not regex.
agent + user + data + destination -
03
Enforce
Allow, warn, redact or block in real time. Redaction swaps values for stable placeholders, so the AI keeps working without the raw data.
allow · warn · redact · block -
04
Prove
Every decision becomes an audit record: which agent, which user, which data, what action and why. Export it for your auditors or your SIEM.
decision log → evidence pack
04 Coverage
One gate across every AI path.
One policy and one audit trail, wherever the AI runs. Here's what we're building with early-access partners now.
Copilotsroadmap
Microsoft 365 Copilot & enterprise copilots
See what copilots pull from SharePoint, OneDrive and Teams into answers, and flag oversharing before it spreads.
Browser AIearly access
ChatGPT, Claude, Gemini in the browser
A managed browser extension checks prompts and uploads before they're sent, and tells personal accounts from company ones.
MCP & agentsearly access
MCP servers & in-house agents
An MCP-aware proxy inspects tool arguments and results, so connectors can't quietly ship records to unknown destinations.
API gatewayearly access
LLM API gateway
Point your apps at Haltgate instead of the model provider. Requests are inspected and redacted, then forwarded. Changing one URL is the whole integration.
05 Compliance
Evidence, not screenshots.
Regulators are starting to ask how AI touches personal data. Haltgate turns each decision into a record you can map to the frameworks you already report on.
GDPR
Data minimization in practice: personal data is redacted before it reaches a model, with a log of which AI processed what.
EU AI Act
Records and human oversight for AI systems. Article 50 transparency rules have applied since Aug 2, 2026, and high-risk Annex III obligations now start Dec 2, 2027. source ↗
HIPAA
Keep PHI such as medical record numbers and patient details out of AI tools you have no agreement with, and show how you did it.
PCI DSS
Card numbers are checked and masked before any AI sees them, which keeps AI tooling out of your cardholder data flows.
Haltgate supports your compliance program. It doesn't make you compliant by itself and isn't legal advice.
06 Why Haltgate
Legacy DLP watches people. We watch what acts for them.
Legacy DLPHaltgate
Built forEmail, endpoints, USB, web uploadsCopilots, connectors, MCP tool calls, agent APIs
Knows the actorThe user, sometimesThe agent and the user it acts for
PolicyHand-written rules and keyword listsPlain policies that weigh data, identity and destination
ResponseAlert, or block the whole actionAllow, warn, redact in place or block
UnderstandsFiles and messagesPrompts, JSON tool arguments, code and configs
RolloutMonths of tuningDesigned to install in a day: an extension, a proxy or a URL change
07 Design-partner program
Build the gate with us.
We're choosing a small group of security teams at regulated companies (financial services, insurance, healthcare and legal) to shape Haltgate before general release.
You get
- A 60 to 90 day pilot at no cost
- A direct line to the founders
- Real influence over the roadmap and policies
- Early-partner terms when you convert
We ask
- A limited or non-production deployment
- Success criteria agreed up front: incidents caught, false positives, time to deploy
- A short feedback call every two weeks
08 FAQ
Straight answers.
Still unsure whether Haltgate fits your stack? Ask us directly when you request access.
Is Haltgate available today?
Haltgate is in early access. We're working with a small number of design partners on the browser extension, the MCP and agent proxy and the LLM API gateway. Microsoft 365 Copilot coverage is on the roadmap.
We already have DLP or Microsoft Purview. Why add this?
Keep them. Haltgate covers the AI layer they weren't designed for: agent tool calls, MCP traffic and prompts built from your data. It's meant to sit next to your current stack and send its decisions to the SIEM you already use.
Does our data pass through Haltgate's cloud?
We're designing inspection to run as close to the data as possible: in the browser for the extension, and in your own environment for the proxy and gateway. Deployment options are something we finalize with each design partner.
Will redaction break our AI workflows?
Redaction replaces each sensitive value with a stable placeholder such as ⟦EMAIL_1⟧, so the model can still reason about the text. Where it's safe, values can be restored in the response on the way back.
Does Haltgate make us GDPR, HIPAA or EU AI Act compliant?
No single tool can. Haltgate enforces controls and produces evidence that supports your program. Your legal and compliance teams stay in charge of interpretation.
What does pricing look like?
Design-partner pilots are free. We'll set commercial pricing together with our partners before general availability.