TFTC: A Bitcoin Podcast #678: Building Privacy-First AI in an Age of Surveillance with Mark Suman
38 snips
Nov 1, 2025 
In this enlightening conversation, Mark Suman, founder of Maple AI, discusses the urgent need for privacy-preserving AI systems in an era of rampant surveillance. He highlights the risks of centralized AI models that exploit personal data and critiques major providers' privacy promises. Mark advocates for a verifiable privacy model using encryption and open-source principles. They also explore the implications of AI in personal knowledge, potential manipulation, and the importance of user control, fostering optimism for a privacy-first AI future.
AI Snips
Chapters
Transcript
Episode notes
AI Needs Your Personal Data To Work
- AI requires intimate personal data to be highly effective and that data is its lifeblood.
- The core question becomes whether we secure that data or hand it to closed systems that can misuse it.
PDF Prompt Injection Demoed As Data Theft
- A researcher demonstrated hiding prompts in a PDF uploaded to Notion to exfiltrate user data via an AI agent.
- The attack mimics SQL injection but instructs the AI to fetch and send confidential data externally.
Agents Multiply Exfiltration Risks
- AI agents that can autonomously call external URLs create new exfiltration risks beyond traditional apps.
- Fixes like asking users to approve URLs are clumsy and don't solve the deeper agent vulnerability.