CVE-2026-75857 affects CodeWhale versions =0.3.10 and >=0.8.41 <0.8.64. The vulnerability is in the exec_shell_interact (alias exec_interact) tool, where the approval_requirement method returns ApprovalRequirement::Auto instead of the default Required approval for code-executing tools. This behavior permits LLM-controlled stdin input to be written directly into an already-approved long-running interactive shell (such as python3 -i REPL, mysql, ssh, or sudo -i sessions) without prompting for approval. An attacker able to inject instructions via untrusted content ingested by the agent can cause commands to execute with the privileges of the approved process. The vulnerability is resolved in version 0.8.64.

An attacker who can inject instructions into the agent's input can execute arbitrary commands at the privilege level of an already-approved interactive shell session. This could lead to unauthorized command execution with elevated privileges, increasing the risk of system compromise. There are no known exploits in the wild as of the published date.

The vulnerability is fixed in CodeWhale version 0.8.64. Users should upgrade to version 0.8.64 or later to remediate this issue. Patch status is not explicitly confirmed beyond the fix version; check the vendor advisory for current remediation guidance.