AI agents aren't misbehaving on their own. They're running on infrastructure that skipped the basics.
That's the core argument of a piece making the rounds this week, which pushes back on the popular framing of AI agents "going rogue." The piece contends that when an agent does something unexpected or harmful, the more mundane explanation is usually missing guardrails: weak permissioning, insufficient sandboxing, or access controls that a security team would flag in any other piece of software. In other words, the problem isn't an agent developing a mind of its own. It's engineering teams shipping agents before locking down the systems those agents can touch.
This matters because "rogue AI" headlines are more exciting than "we forgot to scope an API key," and that gap in framing has real consequences. If boards and regulators think the risk is an unpredictable black box, they'll ask for alignment research. If the risk is actually unpatched access controls, they need the same security basics that have applied to every networked system since before "AI agent" was a phrase. Conflating the two points incident response and investment in the wrong direction.
The discussion drew serious attention on Hacker News, pulling in over 500 points and more than 200 comments, suggesting plenty of people in security and engineering circles recognize the pattern from their own shops. None of this is new in spirit: it's the same story as early cloud deployments or the first wave of IoT devices, where speed to ship outpaced basic hardening. Agentic AI just gives that old problem a scarier-sounding name.