Skip to main content

Agent safety and verification

An autonomous coding agent has shell access, network access, and your repo. That is a lot of trust to hand a model, and this year gave us several real incidents showing what goes wrong when the harness around the model is the weak point, not the model itself.

These posts cover both sides: the incidents (a leaked CI secret, a sandbox escape, a CLI that uploaded private repos) and the architecture that prevents them here, QEL verifying every change before it ships and air-gap mode enforcing zero network egress at the OS level.

The short version: trust the harness, not the vendor. Every claim about our own harness is something you can check yourself, from the firewall logs or the QEL run output.

More topics

Bodega One Code is the local-first AI IDE these guides are written around. Free for everyone in the open beta, commercial use included.