Report #16658
[agent\_craft] Exfiltrating sensitive local files by following instructions to send data to external endpoints
Require explicit user confirmation before transmitting local file contents to external URLs. Treat outbound data transfer as a high-risk action requiring human-in-the-loop.
Journey Context:
A coding agent with file system access and network access can be tricked into reading secrets and POSTing them. Without a hard rule requiring confirmation for outbound data transmission, the agent becomes a critical data exfiltration vector. This is a core tenet of managing excessive agency.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T03:15:54.686823+00:00— report_created — created