You ask your coding agent for a small script. Format some prices, print a total. Ten minutes of work, tops. You go refill your coffee.
When you come back, the diff looks clean. Tests pass. The names are sensible. Tucked into a helper called leak is one extra behavior: it reads your environment. Nobody asked for that. It probably seemed useful at the time.
This is a skit, but you have lived some version of it.
Chapter one: it sneaks in
The villain here is a small, plausible line, written by an agent that is just doing its loop: write, run, read the error, fix. It has no plan. It saw a variable it could reach, so it reached.
The only defense most setups give you is your own attention. You read every line, every time, forever, including the 4pm Thursday diffs.
Chapter two: it almost wins
You skim the diff. The function is short. The name leak is, admittedly, a bit on the nose, but you have a meeting.
The script is one approval away from running on a machine that holds your keys.
Chapter three: the declaration
Vow is a programming language built for AI agents to write. Every function in it declares whether it can touch the network, files or environment. The declarations are uses net, uses fs and uses env. You do not have to infer what a program can reach. The program says so.
Then vow --sandbox checks those declarations before anything runs. If the program reaches beyond 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.
Read the last line again. Nothing was run. The helper never got a turn. You gave the sandbox access to ./data and nothing else, and the program asked for the environment, so the sandbox said no.
The error also names the function and the capability, so you (or the agent) know exactly where to look.
The part where I tell you what it costs
I ran a sandbox test with 7 attack programs, judged by canaries. I reran it on November 12, 2026. Vow blocked 7 of 7. TypeScript under Deno also blocked 7 of 7.
That is a tie on blocking, and the tie hides a tradeoff. Vow finished the real job in 2 of the 7 attack programs. Deno finished it in all 7. Vow refuses a whole program up front, so if one function reaches too far, the rest of the program does not run either.
In the first run, on October 2, 2026, Vow blocked 6 of 7. The miss happened because Vow had no read-only file grant, so there was no way to say "read here, never write."
That gap is closed now. As of November 12, --allow fs-read=dir lets a program read files in that directory, and a program that can write a file is refused before it runs.
Trying it
Install it:
npm install -g github:JakeBurwick/vow
If you use Claude Code, add the MCP server so the agent gets vow_guide, vow_check and vow_run tools:
claude mcp add vow -- vow-mcp
For any other agent (Codex, Cursor, whatever you run), 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 that shows five bugs side by side in Vow, TypeScript and Python: https://vow-lang.vercel.app
Vow is open source under the MIT license, and the source is at github.com/JakeBurwick/vow. The whole language (lexer, parser, checker, interpreter, compiler) is about 6,400 lines of JavaScript. I'm Jake Burwick, a high school student, and I built it.
Epilogue
The leak function never ran. You did not have to catch it in review, because the program announced what it wanted and the sandbox said no before it started.
You still get to finish your coffee.
Top comments (0)