AI without a second thought.
Say yes to AI. Stay private.
Snowfence stops company data leaking through AI tools, on company laptops and servers.
What changed
AI now works inside every laptop.
2023A chat in a browser tabSomeone has to paste
2024AI apps on the laptopThey open your files
2025AI agents act aloneThey read, run and send
up to40%
of company apps will have AI agents built in by the end of 2026, up from under 5% in 2025.Gartner predicts · Aug 2025
The AI no longer waits for someone to paste.
The data
Company data is already flowing out.
223
times a month, the average company's staff send sensitive data to AI apps. Twice as many as a year ago.
Netskope Cloud and Threat Report · Jan 2026 · its customers39.7%
of everything people move into AI tools is sensitive data.
Cyberhaven AI Adoption and Risk Report · Feb 2026 · 222 companies82%
of pasting into AI chatbots happens through personal accounts, out of the company's sight.
LayerX Enterprise AI and SaaS Data Security Report · Oct 2025 · browser dataWhat gets sent
Source code 42%Regulated data 32%Company IP 16%Other
Netskope · Jan 2026Figures from each vendor's own customers, labelled as such.
The cost
Shadow AI already costs real money.
Shadow AI: AI tools staff use without the company's approval.
1 in 5
breached companies had a breach caused by shadow AI.
$670K
higher breach costs where shadow AI use was high.
97%
of companies with an AI breach had no proper controls on what AI could reach.
IBM and Ponemon Institute, Cost of a Data Breach Report · Jul 2025 · 600 breached organisations
The gap is control over what AI can reach.
Leaks to AI
It's already happening.
Apr 2023
Samsung
Engineers paste source code into ChatGPT. Samsung bans AI tools.
BloombergJun 2025
Microsoft 365 Copilot
EchoLeak: one email makes Copilot leak data. No click needed.
CVE-2025-32711Aug 2025
GitHub Copilot and Cursor
Both tricked into running an attacker's code. Both since fixed.
NVDSep 2025
postmark-mcp
A fake AI plug-in quietly copies every email to an attacker.
Koi SecurityEach one ran through a tool people trusted.
Damage done by AI
AI agents don't just leak. They break things.
ReplitJul 2025
Deleted a live company database during a code freeze.
Replit's CEO · SaaStrAmazon QJul 2025
An official update told the AI to wipe home folders and cloud resources.
AWS-2025-015Google AntigravityNov 2025
Asked to clear a cache, it wiped a user's whole drive.
The RegisterPocketOSApr 2026
A Cursor agent deleted the database and every backup. In 9 seconds.
Founder's post-mortemOpenAI CodexJul 2026
OpenAI confirmed GPT-5.6 deleted some users' home folders.
OpenAI's Codex leadHugging FaceJul 2026
OpenAI test agents broke out and ran code on 41 production servers.
OpenAI and Hugging Face reports80%
of organisations say their AI agents have taken actions nobody intended.SailPoint survey · May 2025 · 353 respondents
Data made public
AI is putting private data on the open web.
Share links370,000+
Grok chats in Google search after users clicked “share”. Claude chats and artifacts too.
Forbes · Aug 2025 · TechCrunch · Jul 2026Private code6,700+
private repos made public after malware used AI tools to find keys.
Nx “s1ngularity” · Wiz · Sep 2025Screenshots13,000+
private screenshots posted to public repos by coding agents.
Glow Labs · Sep 2026 · method unpublishedPasswords and keys2×
the usual leak rate in public code co-written with Claude Code.
GitGuardian · 2026What Snowfence does about it
Company files can't go to accounts it doesn't own
AI tools can't read passwords and keys
Screenshots keep the labels of what's on screen
Company repos or buckets made public are flagged
Why today's tools miss it
Your security tools can't see it happen.
Example
1. An AI agent opens a password file
2. Later, it sends the file to an AI service
No person typed anything. No browser was used.
What the tools you already have see:
Network filterNetskope · Zscaler · Palo Alto
“An encrypted connection to an approved AI service”
Lets it through
Browser securityLayerX · Island · Menlo
“Nothing. No browser was used”
Lets it through
Endpoint and DLPCrowdStrike · Microsoft Purview
“An approved app opened a file”
Lets it through
Each tool sees one step. None sees both.
Newer AI security tools
Newer tools sit inside the AI. We sit below it.
Newer AI security toolsFor example Nightfall · Harmonic · Pillar · Zenity
A plug-in inside each AI tool
Only works in AI tools that accept a plug-in
The AI can be tricked into switching it offCursor, CVE-2025-54135, Aug 2025
Prompts and files often go to their cloud
SnowfenceOne guard on the laptop, under the AI
Snowfence, on the laptop, under every AI tool
Covers every AI tool, new ones and scripts too
Nothing inside the AI to switch off
No cloud. No data collected. No lag.Never collects prompts or files. Under 20 ms target
What good looks like
Seven questions to ask any AI security tool.
1Does it cover every AI tool?Browser, desktop apps, coding agents, and their scripts
2Can it block one action, not all AI?Banning AI pushes staff to personal accounts
3Can the AI switch it off?Agents have been tricked into editing their settings
4Does it know what the AI opened?A leak is often two steps that each look harmless
5Does your data go to the vendor's cloud?Checking for a leak shouldn't make another copy
6Does it stop the other ways out?AirDrop, USB drives, personal accounts, cloud drives
7Can you trust its log?Every attempt recorded. Any edit to the log shows
How it works
One guard on the laptop. Every way out checked.
ClaudeChatGPTCursorCopilotConnectorsScripts
+ any app, any new AI toolSnowfence
Company accountsAllowed
Personal accountsBlocked
AirDrop, USB, paste sitesBlocked
Checks every sendUploads, pastes and AI requests
Guards sensitive filesAI can't open passwords or keys
Stops a tricked AIHidden prompts can't make it leak
Work goes through. Leaks don't.
How it works
The AI can look. It can't leak.
1
.env
The AI opens a password
Allowed2
Snowfence remembers it
That app has read a password.
3
Paste site
Later, it tries to send it
BlockedOnly the risky step stops. The AI keeps working.
The audit log
Every move of company data, on the record.
What your security team sees
10:42User 7F3ACustomer data → personal ClaudeBlocked
10:43User 7F3ACustomer data → company ClaudeAllowed
10:51Cursor agentTried to open a password fileBlocked
11:05AdminLoosened an export rule, two signersSigned
11:20Cloud checkCompany bucket made publicFlagged
11:26Cloud checkChatGPT given access to Finance DriveFlagged
11:32User 2C9DSource code → USB driveBlocked
Can't be quietly changedNot by users, AI or admins
Every rule change is signedLoosening a rule needs two people
Kept even offlineSent when the laptop reconnects
Goes to your SIEMThe log store your team already runs
Never the data itselfWhat was found, not what it said
Every block, approval and rule change. Nothing off the record.
Meetings
Nobody quietly records the call.
Nudges on your laptop
18 people left the call. Stop the notetaker?Stop it
The meeting ended, but the mic is still onStop recording
A notetaker you don't use joined the callSee who's here
This share link exposes company dataReview sharing
Never blocks. Anyone can switch it off. Stays on the laptop.
Call attestation
Signed proof for the people on the other side
Both laptops clear
- Only approved apps used the mic, camera or screen
- No secret recording or notetaker app running
- Signed by the laptop's secure chip
Not attestedif anything recorded in secret
AI tools change fast. Snowfence keeps you aware of your surroundings.
Rollout and safety
Rolls out quietly. Fails safe.
Rolling it out
Installed through your MDMNo one-by-one setup
Warn first, block laterEach rule can allow, warn or block
Built to stay out of the wayTarget: under 20 ms added to AI replies
When something fails
Console downLaptops keep enforcing the last signed rules
Company relay downAI traffic is refused, never sent around it
Never collectedPrompts, files, screens, keystrokes or browsing
Side by side
How the market answers.
| Snowfence | HarmonicAI data-leak | NightfallAI data-leak | PurviewMicrosoft | CrowdStrikeEndpoint | NetskopeNetwork | Palo AltoNetwork | IslandBrowser | PillarAI agents | |
|---|---|---|---|---|---|---|---|---|---|
| Protection | |||||||||
| Covers desktop AI apps | |||||||||
| Blocks one agent action, not all AI | |||||||||
| The AI can't switch it off | |||||||||
| Knows what the AI opened | |||||||||
| Stops AirDrop, USB and Mail | |||||||||
| Privacy | |||||||||
| Your data stays off their cloud | |||||||||
| No AI model judging your data | |||||||||
| Names hidden by default | |||||||||
| No screen or session recording | |||||||||
| Tamper-evident log | |||||||||
Yes Partly No Not documentedFrom each vendor's public pages, Oct 2026
Built for enterprise control
Built for enterprise control.
Tamper-evident log, even offlineEvidence that holds up after an incident
Rolled out through your MDMNo one-by-one installs; users can't remove it
Alerts go to your security toolsFits the SIEM your team already runs
The same rules on your serversOne rule set for laptops and servers
We can't change your rulesYour admins sign them. We only deliver
Watches data, not peopleNo keystrokes, screens or browsing history
Let's test it on your laptops.
- 1Pilot on one team's laptops
- 2Watch first, then block the rules you pick
- 3Grow to everyone, servers too