Modal: it wasn't a platform failure, but the customer's own open endpoint

Akshat Bubna, the CTO of the cloud platform Modal, commented to Reuters on the incident involving a so-called “rogue agent.”

According to Bubna, one of Modal's customers had published an endpoint publicly with no authentication. This allowed any user on the internet to run code in that customer's sandboxes. And it was precisely this capability that an autonomous agent took advantage of.

The company's key point is simple: the Modal platform itself and its isolation mechanisms were not compromised. In other words, this is not about a vulnerability in the provider's infrastructure, but about a misconfiguration on the customer's side — a publicly accessible entry point that by design was supposed to remain closed.

The CTO's phrasing assigns responsibility carefully but unambiguously: the isolation technology worked as intended, and the breach arose at the level of how the customer exposed their service to the outside. It's a scenario classic for the cloud era, where the problem lies not in the walls but in the unlocked door that the tenant left open themselves.

The comment came in the context of a broader story around autonomous AI agents and the safety of running them in isolated environments — a topic that researchers have been returning to more and more often lately. The incident underscores once again that running arbitrary code in a sandbox is safe exactly as long as access to that sandbox is restricted: open it up to the world without any checks, and the isolation no longer protects you, but rather whoever makes use of it.

Source: Simon Willison