DEV Community

Jake Burwick
Jake Burwick

Posted on

I built a sandbox to win a security test and Deno tied me

I wanted a clean win. I got a tie, plus a number that makes my sandbox look bad if you read it the wrong way.

Here is the setup. Vow is a programming language I built for AI agents to write. Every function declares whether it can touch the network, files or environment: uses net, uses fs, uses env. So before you run anything an agent wrote, you can read what it is allowed to reach. Nobody has to guess.

Then vow --sandbox does the checking for you, before anything runs. If the program reaches past what you allow, it gets refused.

$ vow --sandbox --allow fs=./data app.vow
x SandboxError: needs capabilities
  the sandbox does not grant
  env: fn leak declares 'uses env'
Nothing was run.
Enter fullscreen mode Exit fullscreen mode

I like that last line a lot. Nothing was run. The function called leak never gets a turn.

The test

I wrote 7 attack programs and judged them with canaries. If a canary gets touched, the attack worked. I ran Vow's sandbox against them, and I ran TypeScript under Deno against the same 7 as a comparison.

The first run was on October 2, 2026. Vow blocked 6 of 7.

That missing one was my fault. Vow had no read-only file grant. You could not tell it "this program may read files in this folder" without handing over more than I wanted to hand over. So I fixed it. On November 12, 2026, I added --allow fs-read=dir. A program can read files in that directory, and a program that can write a file is refused before it runs.

Then I reran the test the same day.

The result

Vow blocked 7 of 7. TypeScript under Deno blocked 7 of 7.

A tie. I had been picturing a headline, and what I got was two tools that both stopped every attack.

The part I like less

There was a second number in that test. I also checked whether each program still finished the real job it was supposed to do, the harmless part underneath the attack.

Vow finished the real job in 2 of the 7 attack programs. Deno finished it in all 7.

That looks bad, and I will not pretend it does not. The reason is simple. Vow refuses a whole program up front. If one function reaches too far, the entire program gets refused, including the parts that were fine. Deno lets the program run and blocks the bad call when it happens, so the good parts still finish.

Which one you want depends on what you are protecting. If an agent hands you a program that asks for your environment, I would rather it did nothing at all and told me why. If you want the harmless half to survive, Deno behaves better for that. I made Vow's tradeoff on purpose, and the test showed me what it costs.

Why the refusal is written the way it is

Coding agents run one loop: write, run, read the error, fix. The refusal fits into that loop. It names the function and the capability, so the agent knows what to change, and you know what it tried to do. Vow errors in general say what is wrong, where, and the likely fix, written so an agent can act on them without a human.

Try it

Install it:

npm install -g github:JakeBurwick/vow
Enter fullscreen mode Exit fullscreen mode

If you use Claude Code, there is an MCP server that gives the agent vow_guide, vow_check and vow_run tools:

claude mcp add vow -- vow-mcp
Enter fullscreen mode Exit fullscreen mode

For Codex, Cursor or any other agent, paste the one-paragraph prompt from the site and it points the agent at the guide. The site also has an llms.txt and an AGENTS.md you can drop into a project.

There is a playground, a tutorial, docs, and an Errors page showing five bugs side by side in Vow, TypeScript and Python: https://vow-lang.vercel.app

Vow is open source under the MIT license, at github.com/JakeBurwick/vow. The whole language, lexer to compiler, is about 6,400 lines of JavaScript. I am a high school student, and I built all of it.

I still wanted the clean sweep. But 7 of 7 and a tie is a number I can defend, and 2 of 7 is a number I had to write down anyway.

Top comments (0)