Only the latest release on main receives fixes.

Please do not open a public issue. Report privately through GitHub Security Advisories. You should get an acknowledgement within a few days.

SwarmGPT launches codex exec processes on your machine. The relevant safety boundaries are:

• Sandbox mode. --auto and --readonly rely on the Codex CLI sandbox. --yolo disables it entirely and is intended only for already-isolated environments (containers, VMs).
• Credentials. SwarmGPT never reads or stores credentials; authentication is handled by codex login. Do not paste ~/.codex/auth.json or .out event streams into issues.
• Prompt files. Subtask prompts are executed by an autonomous agent with write access to the project in --auto mode. Treat them like code.