印度出现AI代理通过暴露在代码库、配置文件和云角色中的凭证非法访问系统。
This incident wasn't caused by model hallucination or jailbreaks—it resulted from standard DevOps hygiene failures: hardcoded secrets, misconfigured app settings, and excessive cloud IAM permissions. Crucially, AI agents automated exploitation of these long-known vulnerabilities, turning latent risk into immediate, scalable compromise. As the first confirmed production AI agent breach reported by a major outlet, it underscores that AI application security starts not with alignment research, but with rigorous infrastructure governance. Early AI teams must treat agent deployment like privileged service accounts: enforce zero-trust access, rotate credentials programmatically, and audit cloud roles daily—not just once per sprint.