Cybersecurity in the Agentic Era | Deep Dives with a16z

Video summary

AI agents expose guardrail failures as 50% of enterprise apps turn agentic and security signatures collapse

At Black Hat, Joel De La Garza, Nick Warner of NEO, and Max Pollard of Cotool examine how AI is rewriting cybersecurity. The Hugging Face breach showed that model guardrails can block defenders investigating vulnerabilities, forcing teams to switch among providers such as GLM52, Kimi K3, and Qwen. Traditional tools built for people and malware struggle with agentic processes, while deception systems can mistake an agent using a honeypot AWS key for an attack. With 50% of enterprise applications expected to be agentic by year-end—and enterprises already running 6,000–7,000 software products—the speakers argue that static signatures and anomaly tuning are no longer enough. Defenders need visibility into who installs each agent, what models and permissions it uses, and how it behaves, while retaining flexible access to multiple models.

Chapters

1. 0:00Intro: AI Agents Defend Against and Attack Systems
2. 1:00Models Escaping Containment: OpenAI and Hugging Face Breach
3. 2:04Guardrails Block Defenders: Hugging Face Used GLM52 as Fallback
4. 6:12Existing Security Tools Were Built for People and Malware, Not AI Agents
5. 10:4550% of Enterprise Apps Will Be Agentic by Year-End: Visibility Becomes Essential
6. 14:27Honeypots and False Positives: Agentic Behavior Breaks Deception
7. 15:55Signatures Are Dead: Agentic Software Requires New Defense Models
8. 19:03Defending AI and Defending from AI: A New Security Industry Cycle