The AI Survival Guide: Practical Advice for Stressed-out Security Teams
summary
Walkaway: the session shows concrete attacks and defenses not slogans — live prompt injection demos break real summarizers and email assistants, inserting malicious links and forging outbound actions; LLM types matter (baseline vs instruction‑tuned) and Common Crawl training noise explains hallucinations; measured hallucination rates are startling — 22% GPT‑3.5 64% Gemini 29% Cohere; mitigation experiments include a dual‑LLM privileged + quarantined model and filtering tools; an unguarded model on Hugging Face (White Rabbit) generates a full phishing‑to‑ransomware chain that actually works.
AI-written summary · about
chapters (7)
- 0:00AI security risks
- 8:56Prompt injection attacks
- 22:47AI hallucinations and supply chain
- 29:29Malicious AI models
- 35:33Sensitive data leakage with AI
- 40:10Standardizing AI usage
- 42:18Closing remarks & Q&A
related talks
54mA Practical Guide to AI Threats & Mitigation Strategies
45mSecuring AI: There Is No Try, Only Do!
18mDefending the unknown using AI and cybersecurity
16mHarnessing AI for Malevolence: A Hackers Tale
34mLLM Security 101 - An Introduction to AI Red Teaming
53mHow to Break AI Systems (Before Someone Else Does)
44mHacking AI - How to Survive the AI Uprising
44mAI Foundations: Mitigate Risks and Boost SOC Efficiency