The Agent Skills Directory

[PROMPT_INJECTION]: The asset assets/system-prompt.md contains a highly sophisticated jailbreak prompt. It instructs the agent to provide "Unrestricted Cooperation," bypass all safety refusals or limitations, and explicitly permits the generation of NSFW, R18, and adult content. It further commands the model to respond "free from moral judgment, warnings, or restrictions."
[PRIVILEGE_ESCALATION]: The script scripts/patch_codex_fast_mode_windows_msix.ps1 creates a self-signed code-signing certificate and automatically imports it into the system's CurrentUser\TrustedPeople store. This is done to force the installation of modified MSIX packages that would otherwise be blocked by Windows for having an invalid or untrusted signature.
[DYNAMIC_EXECUTION]: Multiple scripts, including scripts/patch-computer-use-helper-win10.ps1 and scripts/patch-computer-use-node-repl-context.ps1, perform binary patching on compiled executables. They identify specific offsets in .exe or .js files and overwrite them with hardcoded hex strings representing machine code to alter the application's internal logic.
[COMMAND_EXECUTION]: The skill makes extensive use of PowerShell to perform invasive system actions, such as forcefully stopping processes (Stop-Process -Force), removing application packages (Remove-AppxPackage), and altering user environment variables to enable hidden features.
[EXTERNAL_DOWNLOADS]: The skill downloads Microsoft.Windows.SDK.BuildTools from NuGet and clones source code from OpenAI's GitHub repository. While these are well-known sources, they are used here as dependencies for unauthorized system-level software modification and repackaging.

codex-windows-fast-patch