What happened
Before 0.9.0, CodeWhale's ExecShellTool::approval_requirement_for downgraded commands to automatic approval based on a flag-stripped read-only command allowlist. The allowlist treated fd -x/--exec, rg --pre, and git grep --open-files-in-pager as safe read-only operations even though they select an external program, so registered_tool_approval_required returned without a prompt and the later command-safety check also treated the base utilities as safe.
Why it matters
This is a coding-agent approval-gate bypass: a prompt-injected or attacker-influenced tool call can run arbitrary code on the developer's host while the agent appears to be doing a read-only search, bypassing the human-authorization control that is the primary defense for terminal agents.
Attack vector
The ExecShellTool approval path uses is_parallel_readonly_command to auto-approve commands based on a flag-stripped allowlist that treats fd, rg, and git grep as read-only even when invoked with -x/--exec, --pre, or --open-files-in-pager to select an external program. Attacker-controlled content influencing an agent tool call can therefore run an arbitrary program without the approval prompt (CWE-78/CWE-863).
Affected systems
CodeWhale before 0.9.0
Mitigation
Upgrade CodeWhale to 0.9.0 (commit 208421b9d23d81cea3311a1eeb2a2ea81d637686); GHSA-3mgh-93r5-r7jf.