- Modal's CTO argued against vibe coding the sandbox that AI agents run in, citing reliability, security and scale.
- Our reading is that these failures are hard to see by eye, so buyers should ask who keeps checking the security boundaries.
Akshat Bubna, CTO of Modal, argued on the AI Engineer show that teams should not vibe code the platform their AI agents run in. Vibe coding means letting an AI write software from loose prompts. The platform in question is a sandbox: an isolated workspace where an agent runs code. Bubna gave three reasons: reliability, security and scale.
What was said
Bubna put it bluntly: "I don't think you should Vibe code a sandbox platform." On reliability, he said everyone knows how hard it is to get right, and websites we use daily keep going down. Even with AI, he said, it is hard to reason about what happens when a rare failure, which he called a tail event, breaks the system.
On security, he conceded that you could vibe code a sandbox. His question was about what comes after: who will keep reasoning about the security boundaries and make sure they are defensible while everything changes fast? On scale, he said one customer's personal-assistant workload needs 10,000 sandboxes created every second, with almost a million running, each lasting two or three minutes.
How the boundary works
This part is our plain-language explanation, not his words. A security boundary is the line that limits what an agent's code can reach: files, other systems and the network. Bubna warned that agents can mess up and may leak sensitive data.
The show notes say the talk covers network controls that keep credentials and data from leaking. An agent holding a password and free access to the network could send it somewhere it should not go. One common control is a list of approved web domains. Bubna said "domain-related allow lists are actually not enough." The excerpts we reviewed do not include his reasoning.
Why it matters
Our reading: a vibe-coded app is easy to judge by eye, because you click through it. The failures Bubna described may not show up that way. A rare failure or a weak boundary may never appear in a demo.
For anyone buying or building agent infrastructure, ownership is the question. Who keeps checking the boundary that stops an agent leaking credentials or data? If no one does, the customers whose data the agents handle are the ones exposed. In the part we reviewed, Bubna did not name who pays.
The other side
Modal sells this kind of infrastructure, so Bubna has a commercial interest in the advice. In the portion of the talk we reviewed, roughly 11:45 to 15:46, he gave no example of a vibe-coded sandbox failing and no figures on how often that happens. He did not say AI-written code is always worse. He said the reasoning work around it is hard.
His argument was about whether to build your own platform. How a small team should verify a provider's security is WebPulse's question, not a gap he was addressing.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Don't Vibe Code Your Agent Infrastructure - Akshat Bubna, Modal (2026-11-07)
