The Simple, Yet Lethal, Anatomy of an AI Attack
summary
Shows concrete attacks: a known campaign targets the Ray package and Ray package had no auth, letting attackers leverage dependencies to touch model training; found Hugging Face uploads with leaked tokens and a wrapped model that pickled model opening shell to spawn remote shells; fine-tuned poisoning converts benign LLMs into malware — a poisoned CoPilot injected exfil payload that appends data-leak code to generated lines; researchers harvest millions of Ring videos and used AWS Rekognition to locate passwords, which worked until Amazon patched it.
AI-written summary · about
chapters (9)
- 0:00Introduction to AI security
- 3:39Why latest tech is most hacked
- 8:16Mitre Atlas threat framework
- 11:30Remote code execution in models
- 15:24Poisoned LLM attacks
- 19:09AI as future development language
- 23:06Using AI for classic attacks
- 27:51New era of old threats
- 29:17Q&A and closing remarks
related talks
56mHidden Risks in Open-Source Code and AI Models
52mSleeping With One AI Open: An Introduction to Attacks Against...
51mPrompt Injection: When Hackers Befriend Your AI
36mThe Dark Side of AI: The Hidden Risks in Open-Source AI Models
57mAI as the New Hacker Frontier
37mShadow Vulnerabilities in AI/ML Data Stacks - What You Don’t Know...
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
01h 00mHacking and Securing AI Systems: Playing With Fire and Controlling the Flare of AI