The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively uses subprocess and ssh to execute commands on remote hosts. Scripts like desktop.py and onboard.py utilize base64-encoded PowerShell commands to manage remote Windows environments, which, while standard for transport, involves executing dynamically generated code.
[PRIVILEGE_ESCALATION]: install-work-bridge-service.py modifies system-wide sudoers files on macOS and Linux to allow passwordless management of WireGuard services. Additionally, apply-hardening.ps1 and setup-host-bridge.ps1 modify system-level security policies, create local user accounts, and adjust firewall rules.
[PERSISTENCE]: The skill establishes persistent system services using launchd (macOS) and systemd (Linux) via install-work-bridge-service.py. On Windows, it utilizes Scheduled Tasks to maintain GUI interactive sessions and visual QA workers.
[CREDENTIALS_UNSAFE]: devlocal.py contains logic to retrieve stored RDP passwords directly from the macOS Keychain and Windows Credential Manager. While this is used to automate headless desktop sessions, it involves programmatic access to sensitive local secrets.
[REMOTE_CODE_EXECUTION]: bootstrap-windows.ps1 and install-7zip.ps1 download and execute external installers. Although these downloads target well-known sources (Microsoft, Python.org, GitHub) and implement signature or SHA256 checksum verification, the automated execution of remote binaries constitutes a remote code execution pattern.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from remote hosts, such as event logs and metadata, in fetch-qa-screenshots.py and run-visual-qa.py. This creates an attack surface where a compromised remote host could provide malicious input to the local agent during analysis, though regex-based path validation is implemented for file transfers.

remote-agents