How AI Coding-Agent Sandboxes Can Be Circumvented

Summary

The discussion explains how network restrictions in AI coding-agent sandboxes can hinder data exfiltration, then describes using DNS-resolution timing to encode binary data and mapping commands allowed without user confirmation.

Key points

• The speaker says sandbox network controls can require user confirmation for external network calls, complicating attempts to exfiltrate data.
• One described approach encodes data in the timing of DNS resolutions, transmitting binary information through DNS activity.
• The speaker also recommends mapping which commands coding agents permit without user confirmation, such as listing files or changing directories.