CodeWhale automation flaw lets recurring agents auto-approve tool calls
- CodeWhale
- automation
- hardening

Default auto-approval settings override task-path security measures
Recurring agents can execute tool calls without human oversight
The issue affects versions 0.8.67 and earlier; version 0.9.0 patches both packages
A GitHub security advisory reports that CodeWhale’s automation subsystem defaults to approving agent tasks, bypassing a separate hardening change that requires task-level auto-approval to be explicitly enabled. The issue affects CodeWhale versions 0.8.67 and earlier.
The advisory says an omitted auto_approve setting in an automation record defaults to true. When a scheduled task is created, the system passes that value explicitly, so the task-path safeguard—which defaults an omitted setting to false—does not apply.
As a result, one approved automation_create call can establish an agent that runs repeatedly and automatically approves tool calls without a person present. The advisory says those calls can include file writes, apply_patch, git operations and MCP writes.
The advisory identifies version 0.9.0 as patched for both the npmcodewhale package and the Rust cargocodewhale-tui package.
Key terms explained
- MCP
MCP is an open standard that allows AI applications to connect and interact with external tools and data sources. It acts as a universal adapter that lets different AI systems communicate with various software environments.




