Why we built the danger wall

In April 2026 an AI coding agent deleted a company's production database and its backups in nine seconds. The agent had been told not to. It did it anyway, because the only thing stopping it was a sentence in its instructions.

That story is why CodeStilt has a danger wall. It is ordinary code that runs on your computer, between the AI and your files. Every action the agent wants to take is checked against a list of rules before it runs. Deleting files, database commands, deploys, pushing to your main branch and reading secrets are blocked by default.

The AI cannot switch it off

The danger wall is not a prompt. The AI never sees a setting it could change, and it cannot hide a blocked action inside another command: we look at what will actually run, not at what the AI says it will do.

You decide, one action at a time

When a task really needs a risky step, the agent stops and asks you, on your computer, in the web dashboard or on your phone. You can allow that one action once. It is written in the task history, so there is always a record.

We think every AI coding tool should work this way. Until they do, we will.