You pasted a log
You paste an error to ask what went wrong. Line 40 has your API key in it.
AI guardrails for everyday people & teams
Your secrets stay with you. And we can't see them.
One paste can send an API key, a credit card number or a password to an AI chat, and you can't take it back.
Sight4 checks what you send right on your device, stops keys and credit card numbers before they go, and asks you about passwords. Your prompts never reach us, and we never decrypt your traffic.
AI guardrails for the agents, chats and apps you already use
AI agents
Claude CodeOpenAI CodexGemini CLI: planned
AI chats in your browser
ChatGPTClaudeGeminiGrokKimi
On Chrome today. Safari, Edge and Firefox soon.
AI apps
Cursor
- Finds shadow AI agents and MCP servers
- Catches prompt injection in tool results
- Stops AI account takeover on your device
- Checked on your device
- No root certificate
The problem
API key
Credit card numbers
Tool results
One paste is all it takes.
AI chats and coding tools take whatever you give them, including what you never meant to share. Once it's sent, it's out of your hands.
You dropped in a file
You attach a CSV export to get a quick summary. One column is customer credit card numbers.
A tool handed it over
Your coding agent asks a connected tool for meeting notes. The notes include a key, and the AI reads it next.
Sight4 checks for these on your device, before they're sent. See how it works
Paste without worry
You meant to ask a question, not share your credit card number or the API keys for an app you're building.
Sight4 catches the slip before it's sent and tells you how to fix it, in one step. Here's what works today and what we're building next.
Now Working today, in early access
AI agents and coding apps
Your agent works. Your keys stay home.
The desktop helper plugs into the hooks Claude Code, OpenAI Codex and Cursor already have. It checks each prompt before the agent sends it, warns before risky commands like a force-push, and, once you list the MCP servers you trust, stops the rest.
Works with Claude Code, OpenAI Codex and Cursor. Gemini CLI is planned.
It's a command-line tool for now. A simple installer is on the way.
AI chats in your browser
Ask AI anything. A slip won't cost you.
Keys, card numbers and Social Security numbers are stopped before they're sent, and you're asked before a password or token goes out. You're told what kind of data it was and how to fix it, never the data itself. Text with hidden or invisible characters is stopped or flagged too. If you turn on file checks, attachments are checked as well. You can also choose AI sites to block, or hide Google's AI answers in Search.
Works with ChatGPT, Claude, Gemini, Grok and Kimi on Chrome. Safari, Edge and Firefox are coming soon.
The check runs in your browser. Nothing is sent to us.
Shadow AI
See every AI agent and MCP server on your computer
Sight4 lists the AI apps, coding agents, MCP servers, skills, plugins and AI browser extensions it finds, including the ones nobody remembers installing. It flags keys left in MCP server settings, and skills or rule files such as CLAUDE.md that carry hidden instructions or "run this first" commands.
It runs on your computer and shows you what it found: names, versions, where each one is installed and how it starts, with any keys blanked out. Rule files get counts only, never what's in them, and it never reads your chats. Nothing is sent anywhere. It's a command-line check for now; seeing every device's list in one place comes with the console.
Prompt injection
Planted instructions don't reach your agent
A web page, a document or a connected tool can hide instructions for your AI. Before your agent reads what an MCP tool sent back, and in Claude Code and Codex a command's output too, Sight4 checks it for planted instructions and secrets. In Claude Code, results with strong signs of an attack are held back, and keys in meeting notes are blanked out (from any tool on the Strict setting). Other findings show as warnings.
On Cursor, findings show as warnings for now. What's prompt injection?
AI guardrails for you, your household and your team
One app, one install. What it does for you depends on who you are.
Individuals & Households
Ask AI anything. Keep your card number to yourself.
Plan, write and sort out bills with ChatGPT or Gemini without second-guessing every paste. If a card number slips in, it's stopped before it's sent, and you're asked before a password goes out.
Organizations & small businesses
Say yes to AI at work without reading anyone's prompts.
Your team already uses AI, usually without an IT department. Keys, card numbers and Social Security numbers are stopped on each device before they reach an AI, and Sight4 gives no one, not you and not us, a way to read what your team wrote. One place to set it for everyone comes with the console.
Developers & vibe coders
Let the agent run. Your keys and sign-in files stay off-limits.
Ship with Claude Code, Codex or Cursor knowing your keys won't end up in a prompt. Prompts and MCP tool results are checked where the agent runs, and once you list the MCP servers you trust, the rest are stopped.
Using Sight4 at work? Your prompts are still yours.
For developers and vibe coders
Keep API keys out of Claude Code, Codex and Cursor
Your AI coding tool can read your whole project, including the .env you forgot about. The desktop helper plugs into the hooks Claude Code, OpenAI Codex and Cursor already have and checks each prompt and each MCP tool call before it goes out. In Claude Code and Codex, it also stops the agent from reading SSH keys, cloud credentials and AI tool sign-in files, and warns on .env files. Other files the tool opens on its own aren't checked yet. The model is your provider's; Sight4 is the harness around your coding tool, and its Balanced or Strict setting decides what gets stopped and what gets a warning, and for most checks what happens when a check can't finish.
-
Keys stay out of prompts
API keys, access tokens and private keys are stopped before the tool sends the prompt, with a note on what to remove. Passwords and web tokens (JWTs) get a warning.
-
Stops AI account takeover
AI account takeover is when someone gets the sign-in files of your AI tools and uses them as you. When Claude Code or Codex runs a shell command, including PowerShell on Windows, or Claude Code uses its own file reader or search, Sight4 stops it from opening your SSH keys, cloud credentials, AI tool sign-in files and the Hugging Face token cache. Reading a .env file gets a warning.
-
You pick the MCP servers
List the servers you trust in the settings file, and the rest are stopped. On the Strict setting, you're asked before any server you haven't listed.
-
Tool results get checked too
Before the AI reads what an MCP tool sent back, the desktop helper looks for planted instructions and secrets. In Claude Code it holds back results with strong signs of planted instructions and blanks out keys found in meeting notes; other findings show as warnings. On Cursor everything shows as a warning for now.
-
No root certificate, no traffic interception
The extension and the desktop helper don't sit in your connection to the model provider. Your keys and your traffic stay yours.
Early access is command-line only. Join the waitlist and pick "Developer" to try it.
Privacy and security
We can't see your prompts, by design.
Most companies promise not to look at your data; Sight4 is built so it can't. Your prompts are checked on your device and never reach us.
- Can my boss read what I type into AI?Not through Sight4. Your AI provider's workspace settings, and any other monitoring your employer runs, are separate.
- Does Sight4 see the key it stopped?No. Your device notes only that it was an API key. With the console (coming soon), your admin sees its kind, what Sight4 did, the device and when. Never the key.
Our promises
Your prompts stay on your device
Your prompts are checked on your device. They go only to the AI you send them to, never to Sight4 or your admin.
No one else reads what you write
Sight4 gives no one at Sight4, and no account owner or admin, a way to read what you write.
Your console gets the basics
Your console gets the basics: which devices are covered, counts, and what kind of data was caught. Usage reporting stays off until it's turned on, and your device tells you first.
A look back, just for you
Want to look back? “My activity” keeps your own flagged prompts, encrypted on your device. It's off until you turn it on, and no one else gets a copy.
Every policy change is signed
Every policy change is signed, and your admins can see the log of every signature.
Who sees what, in plain words →
What leaves your device
Today, without a console, there's no account and no telemetry, and the extension makes no network connections of its own.
Sent to us
Nothing.
Stays on your device
- Your prompts, pastes and drops, and the files you attach when file checks are on. They're checked where they are.
- What was found and what happened. A short record of counts, the kind of data, such as "API key", and the decision. In this version the text itself isn't kept.
- Your settings and the detection rules, which are built in, so nothing is downloaded.
- The desktop helper's short log of what it decided, on your computer.
Still goes to the AI site
- Prompts and files you choose to send, as they would without Sight4.
- What you type in the site's prompt box. The site can read it before you press send, even when the send is stopped.
- The site can tell that a Sight4 notice appeared, but not what it says. It can also see what you clicked, and which pasted text was blurred.
Planned: the Sight4 console. When it arrives, this list changes first, before anything new is sent. Your device will then get signed settings from the console and send it device names, versions, which AI tools are installed, usage counts once reporting is turned on, and detection details: its kind, what Sight4 did and when. Never what you wrote. If you turn on My activity, your own flagged prompts are kept here too, encrypted, for you only. The details are in our privacy policy.
Your employer added Sight4. Your prompts are still yours.
- They never get your prompts from Sight4. Your prompts are checked on your device and go only to the AI you send them to.
- They see the basics, never the text. When the console arrives, they'll see which AI tools are installed and what kind of data was stopped, and when. Usage counts are shared only if reporting is turned on for your device, and your device shows a notice before it starts.
- You can check what was sent. While reporting is on, your device shows a sign and lists exactly what was sent.
Your AI provider's workspace settings, and any other monitoring your employer runs, are separate. See who sees what →
Check it yourself
No network code in the extension
In this version, the extension makes no network connections of its own.
See every hook before it's added
aisec hooks install -print shows the exact entries first. Uninstall removes only Sight4's.
A privacy policy in plain words
What we collect, which today is nothing from the product. Read it in a few minutes.
- No root certificate
- No traffic interception
- No kernel driver
- No trackers on this site
A different kind of AI security
Many AI security tools work by collecting what happens in AI tools across a company and analyzing it in the cloud. That's a real choice, with real trade-offs. Here's ours.
| Question | Typical AI monitoring platform | Sight4 |
|---|---|---|
| Where AI use is analyzed | Mostly in a cloud service, after it's collected from devices and AI tools | On each device, before anything is sent |
| Who can read prompt text | Where it's collected, admins with access and the service that stores it | Sight4 never collects it, so no admin and no one at Sight4 can. It goes only to the AI service you send it to. |
| How it's installed | An endpoint agent, a managed browser extension, or a proxy or gateway in front of the AI | A browser extension and the AI tools' own hooks. No root certificate, and nothing that intercepts your connection |
| What the people being protected see | Usually no view of what was collected about them | A notice before any reporting starts, and a list of what was sent (coming with the console) |
| How you buy it | Usually a demo or sales call first | No sales call. Prices go up here when paid plans open. Free for one person on up to 3 devices |
| Who it's built for | Security teams at mid-size and large companies | Everyday people and households, startups, small businesses and organizations |
Plans that fit any size
No sales call, and you buy as many licenses as you need. Paid plans aren't open yet: join the waitlist and we'll tell you when they are.
Free
$0
One person, up to 3 devices.
- Everything that runs on your device: the browser extension and the desktop helper for AI coding tools
- Your own settings
Pro
At launch Price announced when paid plans open
Perfect for a household or a team.
- Everything in Free
- A console to see which devices are covered and set the same protection for everyone
- Usage counts, turned on per device, with a notice on each device first
- Ready-made settings for a household or a team
Organization
At launch Price announced when paid plans open
For startups, small businesses and organizations that need single sign-on, roles and an audit log.
- Everything in Pro
- Groups, with your own keywords and patterns
- Single sign-on and roles
- An audit log, data export and an API
- Run the console in your own cloud at the same price, or keep its signing key there
Each license covers one person and adds 5 devices to a pool you share out however you like. Prices are in US dollars. Pro and Organization start with a 30-day trial for up to 3 licenses, no card needed.
On the roadmap
Coming soonWhat we're building next
The Sight4 console
One place to see which devices are covered and set the same protection on all of them. It will handle settings, counts and what kind of data was caught, never what you wrote.
Tricks typed into the chat
Direct prompt injection
Spotting prompts written to talk an AI out of its own rules, often called jailbreaks.
Hidden instructions in AI replies
Indirect prompt injection
Checking replies in AI chats for instructions planted there. In coding tools, the desktop helper already checks what connected MCP tools send back.
Alerts for AI rule files
A notice as soon as a file your coding tool follows automatically, such as .cursorrules or CLAUDE.md, picks up hidden instructions. Today the shadow AI check scans them when you run it.
Phones and tablets
Bringing the same checks to Safari on iPhone and iPad, plus a Share sheet action and Shortcuts to check text before you paste it into an AI app. Arriving together with Safari on Mac.
What's prompt injection?
It's when someone slips instructions to an AI that you didn't write, to make it do something you didn't ask for, like revealing your information or ignoring its safety rules.
Direct
The instructions are in what's typed or pasted into the chat. For example, a "magic prompt" copied from a forum that tells the AI to ignore its rules.
Indirect
The instructions are hidden in something the AI reads for you: a web page, a document, an email, a reply, or a tool it's connected to. You may never see them, but the AI might follow them.
Questions about Sight4
Can my boss read what I type into AI?
Not through Sight4. Your prompts are checked on your device and never reach Sight4 or your admin. Your AI provider's workspace settings, and any other monitoring your employer runs, are separate.
Does Sight4 see my prompts?
No. Your prompts are checked on your device. Today there's no account and no telemetry, and the extension makes no network connections of its own.
Which AI tools does it work with?
AI agents: Claude Code and OpenAI Codex, with Gemini CLI planned. AI chats in Chrome: ChatGPT, Claude, Gemini, Grok and Kimi. AI apps: Cursor.
Which computers does it run on?
macOS and Windows. The browser extension runs on Chrome today, with Safari, Edge and Firefox coming soon. Safari comes to Mac and iPhone together, through the App Store, and on Mac it needs macOS 12 or later.
Is there a free plan?
Yes. Free covers one person on up to 3 devices, with everything that runs on your device. Paid plans with a console are coming.
What counts as a device?
With the desktop helper installed, a computer is one device, and every browser profile on it is part of that device. With only the browser extension, each browser profile counts on its own. So a laptop with two Chrome profiles and Edge is 3 devices with the extension only, or 1 device with the desktop helper. Each login on a shared computer is its own device.
Join the waitlist
Sight4 is in early access. Tell us who you are and we'll let you know when it's ready for you.