Cogumi AI Shield

Protect against unsafe AI Agents

February 18, 2026 I built an AI agent containment Chrome extension

Hey IH! 👋

I just shipped Cogumi AI Shield, a browser extension that prevents AI agents from causing havoc through uncontrolled tool use. It's been a wild 3-month build, and I wanted to share the journey + get feedback from this community.

What it does:

• Gates agent tool actions (clipboard, network calls, file uploads, form submits)

• Time-bound permissions (e.g., "allow clipboard for 5 minutes" then auto-revoke)

• Content classification: Detects emails, credit cards, SSNs, phone numbers, API keys (OpenAI, Anthropic, AWS, GitHub, Stripe, Slack, Google, SendGrid), and high-entropy tokens in paste operations

• Break-glass pause slider to freeze all agent activity

• Local-first audit log with incident clustering

Why I built it:

I was using ChatGPT for work and realized it could read my entire DOM, access my clipboard, make network requests, and execute actions without me knowing. Agents run with massive permissions and zero user control. Felt like a ticking time bomb for teams handling sensitive data.

The threat model:

AI agents can:

- Read storage, cookies, form fields, clipboard

- Make network requests to exfiltrate data

- Upload files or trigger OAuth flows programmatically

- Execute form submits without user interaction

AI Shield contains these tool abuses through policy enforcement and explicit approval gates.

Technical challenges:

1. CDP bypass prevention — Agents can use Chrome DevTools Protocol to circumvent extensions. Had to use optional debugger permission to detect this.

2. Classification performance — Detecting emails/credit cards/SSNs/phone numbers/API keys in <50ms without blocking UI required careful optimization. Shannon entropy analysis for unknown tokens adds complexity.

3. False positive prevention — Phone numbers vs. dates/IPs, context-aware detection with keyword proximity (±30-50 chars).

4. Grant TTL edge cases — Handling expired grants during async operations was tricky.

5. Policy complexity vs. UX — Too many prompts annoy users, too few miss threats. Still tuning this balance.

Revenue model:

• Free for individuals (always)

• Teams tier coming soon

What I'd love from IH:

1. What features matter?

2. What's the biggest UX friction in security tools you've used?

3. Any blunt feedback on the threat model or positioning?

🔗 Chrome Web Store: [link]

🔗 Website: https://cogumi.ai