One gateway. Three teams. Zero spreadsheets.

In April 2026, attackers worked their way through four of our own AI API keys over five days: CLARA, saskia, MyCloudExpert, and cindyzody-testkey. They were the same kind of keys every engineering team has scattered around: provisioned for a prototype, never rotated, scoped to whatever the model provider's default was.

The keys lived in different repos, different deploy targets, different environment-variable sets. There was no single audit log to consult. There was no hard budget cap that could have stopped the spend at the gateway. There was no inventory that would have told us, on day one, that those four keys had ever been used together. We didn't catch the incident — we caught a credit card alert, and then had to reconstruct what had happened, key by key, from provider dashboards.

We rebuilt the way we wished it had been built. One gateway, every request through it, every call attributed to a project and a Spend Token, every audit row append-only at the SQL layer. We added eight anomaly detectors — including a key_rotation_spike detector that fires when three new keys appear within 48 hours, exactly the pattern we saw in April. We set BASELINE_START_DATEto exclude the breach week from anomaly baselines, so the attacker's usage couldn't poison the model of what "normal" looked like going forward.

The first production client onboarded to the new pattern was UpApply — a cover-letter generation app that had been calling Anthropic directly. We replaced its sk-ant-* with a virtual sk-sr-*key. The real Anthropic credential never touched the application's runtime again. If UpApply's key leaked, it leaked something worthless.

Visionality is the product we wish had existed in April. We're shipping it because we're betting that we're not the only team that found out the hard way.