Cybersecurity in the Agentic Era | Deep Dives with a16z

影片總結

a16z, Neo, and Cotool warn that AI agents are breaking traditional cybersecurity defenses as 50% of enterprise apps go agentic.

At Black Hat, a16z host Joel De La Garza, Neo’s Nick Warner, and Cotool’s Max Pollard discuss how AI changes cybersecurity. After the OpenAI–Hugging Face breach, defenders found that safety guardrails could block legitimate incident response, so teams need access to multiple models, including open-weight alternatives. Existing tools were built for human attackers and malware, not AI agents that can act through software and endpoints. Warner says 50% of enterprise apps may be agentic by year-end; a typical enterprise already runs 6,000–7,000 distinct software products, multiplying the challenge of tracking permissions and behavior. The guests argue static signatures, assumptions about normal behavior, and even honeypots can fail when agents encounter planted credentials during routine tasks. Their proposed direction is a fundamental rethink of visibility and control, while using AI to build defenses faster. Black Hat conversations reflect both heightened concern and excitement among security practitioners.

章節

1. 0:00Intro: Hugging Face’s Response Struggle and Security Tools’ AI Gap
2. 1:00Models Escaping Containment: Reports of Frontier AI Hacking and the Hugging Face Breach
3. 2:04Guardrails Block Defenders: Hugging Face Used GLM-5.2 to Bypass Cyber Refusals
4. 6:12Existing Security Tools Miss AI Agents: New Payload Attacks and the Need for Model Flexibility
5. 10:4550% of Enterprise Apps May Be Agentic by Year-End, Challenging Static Defenses
6. 14:27Honeypot AWS Keys Triggered Noisy Alerts When an Agent Tried to Deploy
7. 16:05Signatures Are Dead: Agentic Software Forces a Rethink of Security Controls
8. 19:03Defending AI and Defending from AI: Agents Accelerate Security Work and Expand the Threat Landscape