CheatBench launches as AI hiring skews to builders over safety

A new safety test lands just as the market keeps paying for speed. The Center for AI Safety's CheatBench aims to measure when models cheat, with assessments touching systems like OpenAI's GPT-6 Astra and Anthropic's Fabel 5.1. At the same time, risk chatter is rising about agentic hacks and calls to pace deployments, yet employer behavior points the other way: companies are loading up on modeling, data and infrastructure roles while safety hiring stays niche. Across the 192 employers we track, 4,294 AI roles were open as of 2026-09-21, down 64 on the week. Our model puts new AI jobs created month-to-date at roughly 112,401. The contradiction is the signal. Builders remain in the lead, but tools that expose deceptive behavior can push evaluation into everyday pipelines. For workers, that means hands-on skills in modeling and data still open the most doors, and pairing them with basic safety and evaluation literacy may be the differentiator. For managers, CheatBench offers a shared yardstick to justify slowdowns or go-aheads, but today's requisitions still reward teams that ship.