[HireRoo, Inc.]What 1.18 Million Real-World Data Points Reveal: Safety and Guardrail Design for AI Agents in Organizations
DAY 2
15:00-
15:30
At CODE BLUE 2026, many attacks targeting AI coding agents and AI assistants will be presented. As one technique after another for bypassing approval prompts and sandboxes is demonstrated, the next question to answer is what your own developers delegate to agents, and where they stop them.
HireRoo measures the actual work logs of developers collaborating with AI agents and has accumulated approximately 1.18 million evaluation records. These logs show how far developers delegated to agents and where they stopped them. In some cases, what appeared to be differences in individual ability turned out to be differences in environment.
In this talk, we present how developers use agents, as revealed by our measurements. We then discuss visualization that helps developers notice their own weaknesses and change how they work with agents, and how to build environments that stop operations that should be stopped through mechanisms rather than human attention. We will share what we have learned from real data on developers who use agents every day.
Speakers:
Kosuke Kuzuoka CEO
-
Location :
-
Track 2(HALL A)
-
-
Category :
-
OpenTalks
-
-
Share :