Virtual talk by Janardhan Kulkarni
summary
Claim: privacy becomes a scaling problem as agents act across many workflows. Approach: train LLMs with reinforcement learning with chain-of-thought using a synthetic dataset of about 700 vignettes, optimize with a policy algorithm (GRPO-style) and reward that balances necessary disclosure versus over-sharing. Result: models trained this way generalize to unseen datasets (PrivacyLens) and materially reduce integrity failures, but remain vulnerable to prompt injection and lack formal guarantees. Built an agent researcher that drafts, implements and audits solutions and even finds bugs.
AI-written summary · about
chapters (10)
- 0:00Privacy scaling challenges
- 1:48Privacy as scaling problem
- 3:41Directions for scaling privacy
- 8:00Behavioral scaling with contextual integrity
- 14:29Evaluating contextual integrity
- 16:00Adversarial threats and limitations
- 20:39Scaling research with agentic systems
- 23:50Provable versus self-improving systems
- 26:20Privacy researcher agent
- 28:29Conclusion and future
related talks
19mTraining Agentic Reasoners
45mPanel Talk
29mBuilding Trust in AI Agents
06mData Privacy Concerns and Solutions When using LLMs in Production
41mBringing stories to life with AI, data streaming and generative agents
19mFrom Legalese to Logic: Translating Privacy Requirements into Verifiable Controls via...
01h 17mPrivacy and New Technologies
30mBuilding Robust, Secure LLM and Agentic AI Workflows