Claude Code hardcoded a Stripe key. Here's how it happens.

8 min · jun 2026 · severity: critical · category: secrets · also at /llms.txt

Ask an agent to “make checkout work” and it will make checkout work. It will read your Stripe docs, write the handler, wire the button — and if your key happens to be in the prompt context, paste it straight into the source. The test passes. The diff looks clean. You ship.

The exact pattern

It almost always looks like this — a working value used directly instead of an environment lookup:

Agents do this because the literal value is the shortest path to a passing test. Nothing in the loop pushes back — the linter is happy, the types check, the demo works. The only thing that would catch it is a gate that runs after the agent and before the deploy.

Why linters don't catch it

Linters check style and syntax, not consequence. A hardcoded string is valid JavaScript. What's needed is a check that knows sk_live_ is a Stripe secret, knows it's in a file that ships to a client bundle, and refuses with a non-zero exit code rather than printing a warning the agent will ignore.

We ran it through the gate

Here's the actual output from scanning the repo above — a real Next.js checkout the agent wrote, before any human review:

fig.1 — one command, ~0.6s, two criticals caught before deploy.

Gate it before the deploy

Add the scan to your pre-push hook or deploy script. Exit codes do the rest: exit 0 ships, exit 1 blocks. Agents understand exit codes natively — which means the same gate that protects you from your agent also works for your agent.

FAQ