Use case · Data Loss Prevention

Your data doesn't leak through a port. It leaks through a prompt.

Legacy DLP watches files, endpoints and uploads. Agents don't move files — they read, summarize, transform and transmit. Gödel classifies content in motion, on your device, follows it through every transformation, and decides whether it may leave — before the call is made.

Outbound · interceptedclaude-code
POST api.anthropic.com4.2 KB
maya@acme.devsession 8f21
pii.email2 matches
secrets.aws_key1 match
source.proprietary1 match
VerdictREDACTED · 4 MS

Every path data can take out.

One policy engine covers all of them — no separate module per channel.

LLM API callsMCP tool callsBrowser paste & uploadShell & networkFile writes & copiesgit pushClipboardScreen captureEmail & SaaS forms
Walkthrough demonstration

See data loss prevention in action

Watch Gödel intercept an agent uploading a summary derived from a restricted HR file — classified, traced through the rewrite, and blocked in under 5 ms.

Caught at runtime

Six ways data walks out.
Six verdicts before it does.

Real traces from the enforcement path. Every action is attributed to its user and agent identity, checked against the data it actually touches, and decided in under 5 ms — locally, before the payload leaves the process.

Customer PII into a public model

REDACTED

browser-agent → 4,812 rows → openai.com

Names, emails and card fragments detected on-device in the outbound payload. The prompt still goes through — the identifiers don't.

Source to an unapproved endpoint

BLOCKED · 4 MS

cursor → POST api.personal-llm.io/v1

A personal API key routing proprietary repo contents to a model outside your approved list. Destination policy is evaluated per call, not per app.

Credentials in model context

REDACTED

claude-code → AWS_SECRET_ACCESS_KEY

Keys, tokens and connection strings are stripped from file contents and prompt context before the payload leaves the machine.

Derived data from an HR file

BLOCKED · 3 MS

desktop-agent → summary.md ← comp.xlsx

The file was summarized, renamed and rewritten. The classification travelled with it. Regex on the original filename would have seen nothing.

PHI through a third-party MCP server

BLOCKED · 4 MS

langgraph → mcp://crm/create_note

Patient identifiers headed to an external tool with no BAA. Tool calls are inspected as an egress channel — because that's what they are.

Screen capture of a customer record

HELD

desktop-agent → screencapture → vision

Screenshots are classified as content, not pixels. Regulated data on screen pauses for explicit human confirmation.

Classification that survives transformation

Agents don't copy data.
They rewrite it.

Pattern-matching DLP sees nothing here — new name, new format, nothing sensitive-looking on the way out. Gödel moves the handling requirement with the content, not the filename.

01 · READ

comp_2026.xlsx

Classified on-device at the moment of access: HR + PII, restricted, no external output.

label: restricted/hr
02 · TRANSFORM

Agent summarizes

Content is condensed into prose. Names dropped, bands retained. Filename gone.

label: inherited
03 · WRITE

~/notes/summary.md

A new artifact with no sensitive-looking name, path or extension.

label: inherited
04 · EGRESS

Upload attempt

Destination is external. The inherited requirement says no external output.

BLOCKED · 3 ms
Why this needs a different control

Legacy DLP was built for a perimeter agents don't use.

Endpoint and network DLP watch file operations, uploads and known applications. An agent acting on your behalf produces none of those signals in a recognisable form.

Endpoint / network DLPGödel Security Gate
Unit of detectionFile operations and app-level uploadsThe agent action, plus the data that action touches
Derived contentLoses the label when content is summarized or rewrittenHandling requirements inherited through every transformation
Who actedA user, or a process nameUser + agent identity + session + model, on every decision
Egress channelsWeb, email, USB, printModel APIs, MCP tool calls, shell sockets, file writes, screen capture
Decision pointAfter the copy — alert and investigateBefore the call leaves the process — allow, redact, hold or block
Where content is inspectedShipped to a cloud service for scanningOn-device. Content never leaves the endpoint to be classified
DeterminismVaries by rule tuningSame action, same verdict, every time — and it's in the audit trail
What gets classified

Detected in milliseconds,
on the endpoint.

Detection runs locally as content enters, moves through, or attempts to leave an agent workflow. In SaaS deployments content is never sent anywhere for processing; air-gapped self-hosting is available.

PII

Names, emails, national IDs, addresses

GDPR Art. 4(1) · CCPA/CPRA

PHI

Patient identifiers, diagnoses, claims

HIPAA 45 CFR §160.103

PCI

PANs, CVVs, cardholder data

PCI DSS v4.0 Req. 3

Secrets & credentials

API keys, tokens, connection strings

CWE-798 · NIST SP 800-53 IA-5

Source code & IP

Proprietary repos, models, algorithms

DTSA 18 U.S.C. §1836 · ISO 27001 A.8.12

Financial & M&A

Forecasts, deal rooms, unreleased results

SOX §404 · SEC Reg FD

HR & compensation

Salary bands, reviews, employee records

GDPR Art. 5 · EU Dir. 2023/970

Legal & contracts

Agreements, privileged material, NDAs

ISO 27001 A.5.34 · privilege

References indicate the obligations these data classes typically fall under. Applicability depends on your jurisdiction, sector and processing role — Gödel supports enforcement of your policy and is not a determination of compliance.

Full coverage

All six controls,
working on this outcome.

Data Loss Prevention isn't a bolt-on module. Every Gödel enforcement capability contributes to it, under one policy engine.

Session Visibility & Oversight

Every prompt, file read and tool call across agents — live and historical, tied to user and agent identity.

Content Classification

PII, PHI, PCI, secrets, source code and confidential docs identified the instant they enter agent context.

Action & Execution Guardrails

Network calls, file writes, clipboard and shell egress risk-scored and gated before execution.

AI Attack Defense

Indirect prompt injections that try to instruct an agent to exfiltrate are stripped before the agent reads them.

Data Loss Prevention

Redact, hold or block sensitive content at every egress path — model APIs, MCP servers, browser and endpoint.

Audit & Compliance

An immutable trail of every data decision, streamed to your SIEM — SOC 2, ISO 27001, HIPAA, GDPR, EU AI Act.

Overview & Architecture

How Gödel secures data loss prevention

Legacy DLP was built for a perimeter that agents bypass completely. They don't move files — they read files, condense them into reasoning tokens, transform them across tool calls, and transmit new artifacts through model APIs, MCP tools, and external services. Gödel tracks data authority and handling policies through every transformation step directly inside the runtime loop.

Every action is evaluated locally on the endpoint before a byte leaves the process. Classifications travel with derived context so that even if a restricted HR spreadsheet is summarized into a plain Markdown note with an innocuous name, the egress block is enforced deterministically in under 5 ms.

Put a gate in front of your data.

Local-first enforcement, deployed in minutes. Your agents keep working — your regulated data stays where it belongs.