Cybersecurity in the Agentic Era | Deep Dives with a16z

影片總結

AI agents break legacy security as Hugging Face faces refusals, 50% of enterprise apps turn agentic

At Black Hat, Joel De La Garza, Nick Warner of Neo, and Max Pollard of Cotool explain why AI agents are forcing cybersecurity to be rebuilt. The Hugging Face breach exposed a defender’s dilemma: guardrails that block attackers can also refuse legitimate vulnerability triage, making access to multiple models—including open-weight alternatives such as GLM52—essential. Traditional tools built for people and malware struggle with agentic behavior, while signatures and even honeypots generate false positives when agents act autonomously. With 50% of enterprise applications expected to be agentic by year-end and companies managing 6,000–7,000 software components, defenders need visibility, endpoint controls, flexible model routing, and a new approach based on what software can actually do—not merely who published it.

章節

1. 0:00Intro: Hugging Face breach exposes defender guardrail problem
2. 1:00Models Escaping Containment: Frontier Labs and Hugging Face breach
3. 2:04Guardrails Block Defenders: GLM52 fallback after cyber refusals
4. 6:12Existing Security Tools: AI agents are neither people nor malware
5. 10:4550% of Enterprise Apps Agentic by Year-End: Visibility becomes essential
6. 14:27Honeypots and False Positives: Agent behavior breaks deception
7. 16:05Signatures Are Dead: Agentic software requires defense from scratch
8. 19:03Defending AI and Defending from AI: Security’s balance of power shifts