Key Takeaways

  • Developers running autonomous AI agents inside standard software containers are operating in an unsafe YOLO mode that ignores basic security boundaries.
  • Docker acknowledged that standard containers cannot safely isolate probabilistic code execution, prompting their shift toward dedicated Docker Sandboxes.
  • Cloud sandbox providers like E2B are pulling ahead because autonomous agents need to spin up and scrap hundreds of isolated environments in seconds.
  • Jerry Murdock warns that model-tool interaction is the make-or-break security layer: if you do not get the sandbox right, the rest of your AI stack collapses.

The Death of YOLO Mode in Agentic AI

Every developer building agentic workflows thinks their setup is protected because they wrapped the model in a Docker container. They are wrong.

Insight Partners co-founder Jerry Murdock points out that the industry is sleepwalking into severe security risks. “I don't know how many developers are running their models in YOLO mode, but probably a lot,” Murdock notes. “They're thinking, 'Oh, I'll put the model in the container. I'll put the tools in the container. I'm okay.' Well, containers aren't safe. You need sandboxes.”

Containers share the host operating system kernel. That design made sense for deterministic web applications, where engineers knew what code was shipping to production. AI agents change the math. When you give an LLM the power to write Python, execute shell commands, and interact with external APIs, it behaves probabilistically. It tests parameters, probes system boundaries, and executes unvetted scripts. A container cannot guarantee containment against an agent running untrusted code on a shared kernel.

Why Docker and E2B Are Rebuilding Isolation

The clearest signal that containers fail agentic security came from container companies themselves.

“This is why the big container company Docker said, 'Hey, containers aren't safe. You better put it in a sandbox,'” Murdock explains. “And that's why they've had a huge success with Docker sandboxes. That's why E2B is successful with cloud sandboxes.”

True sandboxes rely on hardware-level virtualization, microVMs, and isolated execution layers that treat every agent run as hostile. The architecture changes how agents test software. In the container era, spinning up a new environment took seconds or minutes, making disposable testing expensive.

In the agent era, isolation must be instant and disposable. “An agent could say, 'I'm gonna open up a hundred different sandboxes with a hundred different libraries and then determine which is the best app,'” Murdock says. “That knowledge is in a handful of companies today.”

Winning in cybersecurity now requires understanding how models inspect tools and predicting their edge behaviors. As Murdock puts it: “You start there because if you don't get the sandbox right, forget everything else.”

What to Do With This

Audit your agent architecture before Friday. If your application gives an LLM access to a bash terminal, code interpreter, or file system inside a shared container, kill that environment. Replace it with an isolated microVM sandbox using Docker Sandboxes or E2B, and set hard execution timeouts on every tool call.