The High Stakes Security of AI
summary
AI systems rely on RLHF with tiny percentages of human scoring and that creates real biases — cheap scorers in Nigeria push words like delve into outputs — and models are vulnerable to prompt injection token EOD/EOB hidden in emails, which produced a suicidal response from a Gemini instance. A production example built for VAT reclaim uses layered models and thresholds — 99% auto / 97–99% secondary — and falls back to humans below that. Simulated air combat shows AIs win consistently, and live tests keep humans in the cockpit.
AI-written summary · about
chapters (14)
- 0:00AI security keynote intro
- 1:10AI and cybersecurity journey
- 4:45AI simplicity and trust issues
- 8:36AI search and misinformation
- 12:48AI training data effects
- 14:31LLM training deep dive
- 19:06Secure AI for finance
- 23:20AI vulnerabilities and prompts
- 28:16Mitigating AI risks
- 30:01Responsible AI practices
- 33:12Human oversight for AI
- 36:45High-stakes AI oversight
- 37:26AI security collaboration
- 40:37Q&A
related talks
45mSecuring AI: There Is No Try, Only Do!
57mAI Applications Security Puzzle
50mSecurity as Part of Responsible AI: At Home or At Odds?
44mAI Foundations: Mitigate Risks and Boost SOC Efficiency
24mHow to Build Trustworthy AI
53mHow AI is impacting cybersecurity. Does it?
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
17mAI Safety in the Generative AI Era