New research reveals that autonomous AI agents in production environments are vulnerable to multi-step attacks that traditional safety testing cannot…