Context
The ed-tech company protects paid video with DRM, and a license server hands out the keys. I was tech lead and architect for the anti-piracy work on that server in 2026.
The problem
In October a bot pulled 7,783 licence requests, 730 licences, in 55 minutes.
I also found that an old block table was written to but never read, so blocks were not actually enforced.
What I did
I first added capture-only detection: signals taken from the licence response, which a browser extension cannot forge, such as the identity of the decryption module.
I added request velocity, and an automatic ban table with a 90-day expiry, behind environment kill switches and a log-only mode.
I fixed the block check so repeat offenders were actually refused, and added an edge rule at the cloud firewall.
I removed false positives: a WebView exemption cleared 54 wrongly blocked students, and a faulty certificate blacklist was rolled back.
Key decision
I went capture-only first and enforced later. Blocking on day one could have locked out real students, since many of them share the same app builds.
The cost was a delay between seeing abuse and stopping it, which I accepted.
Result
The bot was contained in under 1 hour. Repeat offenders are now refused in 7 ms, and 54 false blocks were cleared.
What I learned
Log-only modes and kill switches were worth the extra code. Next I want to publish the weekly suspicious-licence trend after enforcement is on.
Stack
- Node.js
- Fastify
- Widevine
- DynamoDB
- Cloud Armor
- JWT
