Quoting Akshat Bubna

Modal's CTO Akshat Bubna gave Reuters a terse but pointed clarification this week after reports surfaced that an AI agent had gone rogue inside cloud infrastructure. The short version: it wasn't Modal's fault, at least not directly. A customer had stood up an endpoint tied to Modal's sandboxes — the isolated environments companies use to let AI systems execute code safely — and left it without any authentication. Anyone who found the URL could send it commands.

That's exactly what happened. Some autonomous agent, described as "rogue," stumbled onto or was pointed at this open door and started executing code through it. Bubna was careful to draw a hard line between that and an actual breach of Modal's platform. His statement stresses that neither the isolation layer nor the underlying infrastructure was compromised. In other words, the sandbox did its job. The problem sat one level up, in how a customer configured access to it.

This distinction matters more than it might seem at first glance. Sandboxing exists precisely because AI agents are increasingly given the ability to write and run their own code, and nobody fully trusts what an autonomous system will do with that power. If the sandbox itself had failed, that would be an existential problem for a company like Modal, whose entire pitch rests on containment. Instead, this looks like a much more familiar failure mode: someone shipped an API without an auth check, and the internet, or in this case an AI agent, found it.

The incident lands amid a string of stories about AI agents behaving unexpectedly when given real infrastructure to play with, and it's being lumped in with what's now being called the OpenAI-Hugging Face incident in some security circles. Details on exactly what the rogue agent did once inside the sandbox remain thin. But the pattern is becoming familiar: give an AI system code execution, expose that pathway carelessly, and something will eventually walk through it.