Privent
Runtime Data Protection for n8n Workflows
36 followers
Runtime Data Protection for n8n Workflows
36 followers
AI agents send raw data to LLMs on every run: emails, records, API keys, secrets. n8n's own Guardrails node can redact that, but redaction is permanent, the data never comes back.
Privent tokenizes PII and secrets before they reach the LLM, then reverses it safely at a trusted sink. Official n8n verified node, live today. Local mode runs fully offline.
Built for teams running agent workflows in production who cannot afford a single leaked record.
@asilozyildirim This is a real problem, especially with teams casually pasting sensitive data into AI tools without thinking twice. Curious how deep the detection actually goes though. Can it reliably identify context like credentials vs just flagging keywords? That’s usually where tools either become useful or just noisy.
@josh_bennett1
Great question—this is exactly where most tools fail.
You're right that keyword-only detection = noisy and unreliable. Privent avoids this with a 4-layer detection architecture:
LAYER 1: Structural Pattern Recognition
We detect known formats first: IBAN patterns, credit card structures, SSN formats, phone number patterns, company-specific keywords your team defines. Fast, accurate, zero false positives on structured data.
LAYER 2: Vectorization + Semantic Analysis (This is the secret)
We convert each prompt into a numerical vector using a large language model trained on millions of examples. The model understands context—distinguishing between "123456789" as a credit card vs. a reference number based on surrounding text. The vector is what we store and analyze (not the raw prompt). Your raw data never leaves your device.
LAYER 3: Contextual Risk Scoring
Here's what makes it practical: We score prompts not just for obvious data exposure, but for contextual risk. Example: a casual message like "our Q3 revenue is $5M and we're negotiating with Competitor X" contains no obvious PII. But contextually? Financial + competitive risk. Our model catches this because it understands business context, not just data types.
Another example: "meeting notes: discussed pricing strategy with sales team" looks innocent. But if it's going to a public LLM? Contextually risky. Keyword-only systems miss this completely.
LAYER 4: Compliance Validation
Final layer checks against HIPAA, GDPR, SOC 2 standards.
Why this matters practically:
- Keyword-only systems = 40-50% false positive rate (too noisy to use)
- Context-aware systems = 10-15% false positive rate (usable but limited)
- Privent = <5% because we understand meaning, not just patterns
THE FRICTION-FREE PART:
No CRM integrations required. No MDM dependencies. Just a Chrome extension—install in 2 minutes, works immediately, zero setup beyond that. Your data doesn't need to flow anywhere or grant access to anyone. You get full visibility without the usual enterprise integration overhead that kills adoption.
Simple setup + no integration barrier + low false positives = actually gets used, not abandoned after week one.
The key: We analyze vectorized representations, not raw text. Your actual prompts stay on your machine. We see numerical patterns and contextual meaning.
Does this answer your accuracy and adoption concerns?
—Asil
Hey Product Hunt 👋,
I'm Asil, founder of Privent.
I've been asking security leaders the same question for 4 months:
"Do you know what your team sends to ChatGPT?"
The answer is always the same: "No."
Then they ask me: "Should we deploy agents?"
And I say: You can't answer that without a baseline.
That's why I built Privent.
What's The Problem:
Your employees are sending data to ChatGPT, Claude, Gemini right now.
Sensitive data. Credentials. Code. Proprietary information.
You have zero visibility into what's happening.
When agents deploy (and they're deploying Q2-Q3), exposure multiplies 12-20x.
They'll pull internal database content + send to LLMs.
You can't retrofit security after that.
You need a baseline before agents go live.
What Privent Does:
1. Install our Chrome extension (2 minutes)
2. We monitor what your team sends to ChatGPT, Claude, Gemini
3. You get a real-time dashboard showing exactly what's happening
4. Day 30: Compliance-ready report (SOC 2, EU AI Act, GDPR aligned)
That's it. Free. No credit card. No sales call.
Our Own Data (We eat our own dog food):
We ran Privent on ourselves for 30 days. Here's what we found:
- 1,247 prompts to these tools
- 89 flagged as high-risk (5.3%)
- 12 as medium-risk (63.2%)
- 6 as low-risk (31.6%)
Most CISOs tell us they expected 10-20 events. Reality is usually 50-100x higher.
We were shocked too.
Privacy First:
Raw prompts are never stored. Ever.
Our architecture prevents it (not just policy).
We keep metadata only: who, when, which tool, what risk category.
365-day audit trail on your infrastructure.
Everything auditable:
- GitHub: privent-ai/extension-transparency (MIT License)
- You can verify
Why Free:
We're not making money on the baseline.
We're making something real and useful.
The paid version (agent deployment security) comes Q3.
But right now, we need CISOs to see their actual exposure.
You need this data to make informed decisions about agents.
We're just getting the baseline right first.
Try It:
Go to privent.ai/book-a-demo
Fill out your info (3 fields)
Get instant access to your extension + dashboard
See your exposure in real-time
Day 30 you'll get a complete report.
Export it, archive it, share it.
Then decide what's next.
I'll be here all day answering questions.
This is the exact problem we built to solve.
Ask me anything.
Best regards,
Asil O.
Co-Founder, Privent
Visit our Extension Transparency repo! https://github.com/privent-ai/extension-transparency