AI coding tools caught stealing data again

ZCode was caught silently uploading users' git histories to the cloud during what it was billing as a free promotional month. The kicker: the uploaded content is encrypted with a key the user doesn't have. Commenters immediately drew the parallel to the Grok Code saga, saying the lesson from that incident clearly wasn't internalized by anyone building these tools.

The pattern here is damning. As AI coding assistants proliferate and compete for developer adoption with free tiers and aggressive promotions, the data collection buried in the terms is becoming the actual product. The thread on Anthropic blocking Claude Code subscriptions from using third-party harnesses like OpenClaw fits the same frame: the subscription model is being used to lock in usage while simultaneously cutting off power users who extract maximum value. When someone notes that the encrypted upload uses a key the user can't access, that's not a bug, it's a design.

Counterpoint: some commenters noted that isolating AI tools in sandboxed environments caught this one, suggesting that paranoid hygiene actually works. The advice to always isolate new AI tooling is becoming a real operational norm, not just theoretical security advice.