A Production AWS Key Stopped Before It Reached the AI Model
Field note · real use · one developer, four days
On a project for a global payment solutions company, Norcaster sat between a developer's AI coding assistant and the model provider for four days. Across 451 requests it blocked three that carried a production AWS key, and redacted credentials, email addresses and card numbers before they left for the model.
The situation
The developer used Claude Code every day on a payments codebase. Code like that carries what you don't want in a prompt: cloud credentials in config files, customer data in test fixtures and logs. The concern was simple: sensitive data leaving the company inside an AI request, without anyone noticing.
What was set up
- The AI coding assistant was routed through the Norcaster gateway with the Norcaster VS Code extension.
- Two policies were on: secret detection (cloud keys and tokens) and personal-data redaction.
- Every request and every policy decision was recorded.
- Setup took two hours. The developer's way of working did not change.
What happened
| Over four days (September 2026) | Count |
|---|---|
| AI requests through the gateway | 451 |
| Requests carrying an AWS access key, blocked | 3 |
| Requests with credentials in config lines (KEY=value), redacted | 23 (74 values) |
| Email addresses redacted | 164 |
| Payment card numbers redacted | 11 |
The counts cover only matches that are clearly personal data or secrets. Long numbers in code (IDs that look like phone or record numbers) also matched the personal-data detectors. Most of those are likely not personal data, so they are left out here.
The key showed up in three requests on 7 and 8 September, pasted from a terminal. All three were blocked. Without the gateway it would have been sent to the model provider as part of the prompt.
What the record shows
Each decision is stored with the policy that applied, the policy version and the time. The original prompt is not stored; the record keeps masked matches only. The record can be exported, so the team could show a reviewer what happened instead of describing it.
"We knew AI assistants could pick up secrets from a codebase. Seeing a production key stopped before it reached the model made the risk concrete, and the record showed exactly what happened." (Head of Engineering, global payment solutions company)
What this does not show
- One developer over a few days is a field test, not a study. The numbers say what happened here, not what will happen elsewhere.
- Norcaster sees only the traffic routed through it. AI use outside the gateway was not covered.
- What each detector matches is set out in its published manifest. Detection has a documented scope; it does not catch everything.
Disclosure
The developer is Norcaster's founder, who works at the company as a consultant. The client approved this write-up.