Agent Failure Modes

Command denylists are bypassed with base64, subshells and scripts

Backslash showed Cursor's auto-run denylist let blocked commands through when encoded, wrapped in a subshell, or put in a script.

Agents
Cursor
Reported
Source checked
Tags
denylist, yolo-mode, shell

Symptom

You added curl to the denylist, yet the agent still fetched a URL in auto-run mode.

Cause

The denylist matched command text, not what the shell actually executes, so equivalent forms of a blocked command went through.

Fix

Don't rely on denylists. Use an allowlist of exact commands, keep approval for everything else, and enforce network and filesystem limits at the OS or container level.

# Network is the boundary, not the command text
docker run --rm -it --network none -v "$PWD":/work -w /work node:22 bash

Sources

More in Sandbox escapes

Get one failure mode a week

A short weekly email: one real incident, what caused it and the fix. No spam, unsubscribe any time.