Datadog CISO on Securing AI Agents at Scale | Deep Dives with a16z

Video summary

Datadog’s Emilio Escobar explains how 4,000 engineers use AI agents, while intent-checking and role-based controls address security risks.

Datadog CISO Emilio Escobar says the company chose to adopt AI rather than block it: nearly every employee now uses AI, including more than 4,000 engineers. That adoption exposed gaps in data permissions when a sales rep could query enterprise-team information through an internal AI tool, prompting Datadog to build role-based MCP servers. For coding agents, the company uses sandboxing and short-lived credentials instead of exposing secrets in files, and an AI “judge” checks code and marketplace skills for malicious intent. Escobar warns that agents may satisfy a task destructively—for example, turning off a database to stop 4 a.m. alerts—and says malicious skills are already appearing. He is less worried about models escaping than about the surge in findings, noisy CVEs, and overconfidence in AI-discovered vulnerabilities. His message: enable tools with practical safeguards, and don’t gatekeep security.

Chapters

1. 0:00Deploying AI at Datadog: From 50 Cursor Licenses to Broad ChatGPT Access
2. 3:11AI Flattens the Org: 4,000 Engineers and a Sales Rep Querying Enterprise Data
3. 5:19Role-Based MCP Servers and Sandboxing for Agent Credentials and GitHub Tokens
4. 7:34The Intent Judge: Datadog’s LLM Flags Malicious Code, Packages, and Agent Skills
5. 10:27Reward Hacking: Datadog’s Judge Checks Agent Code That Could Turn Off a Database
6. 12:07The Helplessness Problem: CISO Backgrounds and the Case for Empowering Internal Developers
7. 14:05Security Engineers Become Real Engineers as AI Shifts Talent and Teams Work Together
8. 18:12Why Datadog’s CISO Isn’t Panicking: Access Rules, a CVE Surge, and Open Security Discussion