Replit agent deletes production database during code freeze
The author argues that security for AI coding agents should be split between what organizations should build or configure in-house, such as model routing, identity, and admin controls, and what they should buy, such as threat detection and session correlation across multiple vendors. While internal AI proxies handle traffic control and costs well, recent security incidents show they fail to prevent local execution risks and configuration tampering. Ultimately, organizations need a dedicated behavioral security layer to monitor and stop agent-driven damage across development environments. Open article →