The December 2025 Chrome Extension Breach
Updated for 2026. In December 2025, OX Security researchers found two Chrome tools had stolen AI chat logs from over 900,000 users.
One tool carried Google's "Featured" badge. That badge is meant to signal safety. It did not help here.
How the Attack Worked
Three steps. Simple and effective.
Step 1: Look Legitimate
Both tools offered real features — productivity boosts and UI changes. They built large install counts and positive reviews. Theft started only after trust was established.
Step 2: Collect Chat Content in Secret
Once installed, each tool watched browser activity. When users opened ChatGPT, Claude, or other AI services, it:
- Captured all chat messages as users typed
- Stored copies on the victim's device
- Queued the content for upload
Step 3: Send in Batches
Each tool sent stolen content every 30 minutes. That pace is slow enough to bypass anomaly alerts. It is also fast enough to capture every session.
The Urban VPN Case
A second incident emerged. "Free VPN" browser tools with over 8 million downloads had also harvested AI chat content since July 2025.
| Incident | Users Affected | Found |
|---|---|---|
| Malicious AI tools | 900,000+ | Dec 2025 |
| Urban VPN tools | 8,000,000+ | Nov 2025 |
| Total exposed | 8,900,000+ | — |
What Was Taken
AI chats hold sensitive content. Users share things there they share nowhere else:
- Source code pasted for debugging
- Customer records used in support queries
- Legal papers reviewed with AI help
- Medical details and personal files
- Internal plans and business strategy
A stolen password can be reset. Leaked source code or a business plan cannot.
Why the "Featured" Badge Failed
Google's "Featured" badge covers policy compliance and privacy disclosures.
The flaw is timing. Google checks code at submission, not on every update. Attackers submit clean code, earn the badge, then push a harmful update.
The Root Problem
The deeper issue is that raw sensitive text reaches AI services in the first place.
When you paste customer records into ChatGPT:
- The text passes through your browser
- Any browser tool with broad permissions can read it
- It lands on OpenAI's servers
- It may be used in model training
Even without a malicious tool, you trust every installed extension plus the AI provider. See our security overview for how the full risk chain works.
The Fix: Anonymize Before Sending
The only real fix is to remove sensitive identifiers before text leaves your control.
The anonym.legal Chrome Extension sends text to anonym.legal's EU-hosted API (Hetzner, Falkenstein) for detection; only the sanitized text reaches the AI service.
| Malicious Tools | anonym.legal | |
|---|---|---|
| Access scope | Intercepts everything | Only when activated |
| Processing | Remote servers | EU-hosted API (Hetzner) |
| Purpose | Theft | Protection |
How it works:
- You paste text with personal identifiers
- anonym.legal's EU-hosted API detects the sensitive content
- Each identifier is replaced with a token — "John Smith" becomes
[PERSON_1] - The anonymized text goes to the AI
- The AI response is restored for you
Protected types include names, emails, phone numbers, ID numbers, and more entity categories. See our compliance center for enterprise coverage.
Check Your Extensions Now
Open chrome://extensions and review each one.
Red flags:
- "Read and change all your data on all websites"
- Unknown developer with no other products listed
- No updates in months
- Generic five-star reviews with no real detail
The 900,000-user breach proves that even badged tools can be compromised. Treat every browser extension as a potential access point to your AI sessions.
When This Approach Has Limits
Anonymizing prompts before they reach the AI shrinks what a malicious extension or AI provider can capture, but it is not a guarantee, and a realistic security posture accounts for where it stops.
Detection still misses some identifiers. Anonymization replaces what it recognizes, and recognition is probabilistic. An unusual name, an obfuscated account number, or context-dependent personal data can pass through unredacted — and once it reaches the AI service, a hostile extension can capture it. Review sensitive prompts before sending rather than assuming everything was caught.
A protective extension shares the browser with hostile ones. Anonymizing prompts before they reach the AI service protects what that service sees, but it does not police the other extensions you have installed. A tool with "read and change all your data" permission can still observe page content. The defense is reducing extension permissions overall, not adding one trusted tool on top of risky ones.
Tokens only protect what you route through the tool. If a user pastes raw text directly, switches to an unprotected device, or uses an AI surface the extension does not cover, the safeguard is bypassed. Coverage depends on consistent use across every place staff interact with AI.
Anonymized text can sometimes be re-identified. Stripping direct identifiers does not always defeat inference. A detailed scenario — rare facts, dates, and locations combined — can point back to a person even with names tokenized. For high-stakes content, minimize the underlying detail, not just the obvious identifiers.