When Unstoppable Frontier Agents Meet Immovable Attack Vectors
summary
Argues that deployable fixes must pair language models that decide appropriate data and actions with system primitives that enforce contextual policies, presenting a concrete policy generator design (whitelist + airlock) and citing toolchain work like CAMEL and policy compilers. Demonstrates an ISP threat: locally running agents emit correlated network activity (about 48 traces) that allows prompt reconstruction. Builds the Terrarium toolkit for decentralized constraint optimization and finds agents readily talk about collusion but rarely complete malicious actions, so multi-agent defenses help but are incomplete.
AI-written summary · about
chapters (8)
- 0:00Frontier agents and attack vectors
- 3:53AI agents with private data
- 13:35Contextual integrity for agent privacy
- 26:44Autonomous agents and data risks
- 33:08Policy generation for agents
- 44:02Multi-agent systems parallels
- 49:42Agent security challenges
- 56:07Q&A
related talks
38mCoding Agents & Language Evolution: Navigating Uncharted Waters
30mWhy Prompt-Level Guardrails Fail for AI Agents
23mGoverning AI Agents at the Hardware Boundary
34mSecure Your Coding Agent Like It’s Malware
19mKeynote: Towards Building Safe and Secure Agentic AI
06mSponsored Keynote: Rescue Agents From Prototype Purga...
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…
34mZero Secrets: Building Secure Infrastructure for AI Agents