Security Implications of Artificial Intelligence Synthesizers
summary
Generative models like GPT-4 use much larger context windows (about 32,000-token context) than GPT-3 (2048 tokens), enabling longer coherent output but still producing hallucinations and buggy code; adversaries exploit training pipelines via adversarial data poisoning and users exploit interfaces via prompt injection and prompt exfiltration (Bing leaks multi-page instructions, leaked API keys). Image generators spark copyright suits (Stable Diffusion vs Getty). Studies and tools show detection is unreliable; GitHub CoPilot research finds roughly 40% vulnerable code in early evaluations.
AI-written summary · about
chapters (12)
- 0:00RSAConference 2023
- 1:55Disclaimer
- 3:39Artificial intelligence
- 9:31What can Al train on?
- 11:24Adversarial data poisoning
- 12:36Prompt injection
- 15:27Prompt exfiltration (leaking)
- 16:29Phishing
- 17:39Code synthesis
- 20:43Security operations examples
- 21:36Detecting synthesized content
- 24:40Summary
related talks
30mBattling the Machine: Cutting-Edge Attacks and Defenses in Generative AI
31mHacking AI Applications
55mAI Showdown: The Good, The Glitchy, and The Risky
29mThe Rise of AI: A Reality Check on Generative AI
35mGenerative AI in the real world
55mUsing GenAI on and inside your code, what could possibly go wrong?
52mUsing GenAI on and inside your code, what could possibly go wrong?
27mSafeguarding Generative AI Applications: Navigating Security Challenges and Best Pract...