Pennywise - Invisible privacy risk in the new AI era
summary
Demonstrates concrete privacy bypasses against ChatGPT, Bard and Bing using prompt-injection and vision tricks: a lyrics prompt hack coerces models to reveal contact details, and an image necklace extraction reads personal data from photos; also shows an old LangChain CVE that executes arbitrary code in prompts. Finds inconsistent safety baselines across versions, apps falsely claim no third‑party sharing, and notes Italy’s temporary ChatGPT ban and OpenAI’s subsequent data‑retention and transparency changes.
AI-written summary · about
chapters (11)
- 0:00Invisible privacy risks in AI
- 1:06LLM privacy threats overview
- 9:33Prompt injection deep dive
- 16:00Bypassing LLM safeguards
- 19:14Legal perspective on AI privacy
- 23:41EU regulations vs OpenAI
- 28:46AI provider obligations
- 35:22User rights in AI systems
- 37:49Privacy by design for AI
- 38:53Future of AI privacy governance
- 41:00Q&A
related talks
43mWords as weapons: The dark arts of Prompt Engineering
53mHow AI is impacting cybersecurity. Does it?
18mDefending the unknown using AI and cybersecurity
53mHow to Break AI Systems (Before Someone Else Does)
51mPrompt Injection: When Hackers Befriend Your AI
37mThe Shadows That Follow AI Generative Models
31mHacking AI Applications
25mPreventing an OWASP Top 10 in the world of AI