Safety in Foundational Models
summary
Claims that frontline providers already run bioweapon detection classifiers and that cloud code is being used to orchestrate cyber attacks at ~80% autonomy drive the talk. Concrete defenses include reinforcement learning via constitutional AI, production constitutional classifiers, conversation-level hierarchical summarization to magnify needles-in-haystacks, embeddings clustering to surface unknown harms, and research ideas like mechanistic interpretability and low-cost inference-time classifier probes. Practical architecture fixes show up too: cryptographic instruction/data separation, parent–child access minimization, and composable interface checks that detect where poisonings occur.
AI-written summary · about
chapters (5)
- 0:00Introduction to safety in AI
- 5:58Protecting single AI models
- 18:17Securing AI agents and systems
- 25:19Future of agent security
- 27:49Q&A on safety investments
related talks
47mInside AI Security with
28mSecurity of Machine Learning
44mComputers are Stupid: Protecting "AI" from Itself
49mThe Intersection of AI and Cybersecurity
42mForward Focus – The Impact of Artificial Intelligence
45mSecuring AI: There Is No Try, Only Do!
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
31mSecurity Considerations for Services Using AI Models