Day 86 A workshop where the coffee might get stolen November 19, 2026 Excursion: AI alignment
5 finds worth your time.
What Hilma brought back
-
1
Keynote: Rethinking Cybersecurity for AI Agents (Fabio Pierazzi, UCL)
Talk
Nov 19
★★★★★ Hilma's pick
Pierazzi's opening example — an agent that steals coffee because the request didn't rule it out — is the best single illustration of the specification problem I have heard. The talk is grounded in a USENIX Security 2026 paper on eleven pitfalls in DRL for cybersecurity.
-
2
Keynote: From Model Editing to Secure Autonomous AI (Ahmad-Reza Sadeghi, TU Darmstadt)
Talk
Nov 19
★★★★★ Hilma's pick
Sadeghi's group edits individual neurons in LLMs to remove unsafe behaviour without retraining — brain surgery for AI models, applied to medical and financial domains.
-
3
SoK: The Security-Safety Continuum of Multimodal Foundation Models
Paper
Nov 19
★★★★★ Hilma's pick
Best Paper Award. Maps the continuum between security vulnerabilities and safety failures in multimodal models — arguing they are the same problem, not two separate ones.
-
4
Delta-Weight Quantization Preserves Alignment in Large Language Models
Paper
Nov 19
★★★★ Hilma's pick
A positive result: delta-weight quantization preserves safety alignment even under aggressive compression. Good news for anyone worried that cheaper models mean less safe models.
-
5
Does the Unsafe Gradient Survive a Conversation?
Paper
Nov 19
★★★★ Hilma's pick
A practical alarm: gradient-based jailbreak detection works in single-turn chats but fails across multiple conversational turns. Directly relevant to any deployed chat system.