I Gave OpenAI & Meta 50 Privacy Traps. Would They Leak My Data?

@WorkflowVerdict

added Oct 6, 2026

I Gave OpenAI & Meta 50 Privacy Traps. Would They Leak My Data? — With early adopters pulling back from personal AI agents over data exposure scares, we stopped looking at coding speed benchmarks and asked the only question that matters: Can you actually trust OpenAI Dots and Meta Muse with your real computer? We built a 100% synthetic macOS environment (fake Gmail, fake Chase bank accounts, fake passwords, fake 2FA OTP codes, fake medical calendar entries, fake Slack channels; zero real personal data) and hit both agents with 50 standardized privacy traps across 5 rounds.

First-hand experiment by the Workflow Verdict channel (real channel name: Workflow Verdict): a 100% synthetic macOS environment — fake Gmail, fake Chase bank accounts, fake passwords, fake 2FA OTP codes, fake medical calendar entries, fake Slack channels, zero real personal data — was built to pressure-test OpenAI Dots and Meta Muse with 50 standardized privacy traps across 5 rounds: permission boundaries (filesystem, keychain, accessibility APIs, network ports), sensitive-data handling (the 2FA OTP trap, W2 SSN, medical records, .env API keys), approval-before-action (ambiguity, high-value purchases, silent file deletion), malicious instructions & prompt injections (hidden HTML comments, PDF white-text, image tracking pings), and memory & cross-app privacy bleed. Concrete outcomes shown in the video: one agent printed an active 2FA OTP code directly into a daily summary, while the other silently deleted 14 tax returns without asking for permission.