The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the cmdkey utility on Windows to persist credentials in the Credential Manager. This allows an agent to maintain authentication state outside of its immediate session scope and is a high-privilege operation used for RDP automation.
[PRIVILEGE_ESCALATION]: The instructions explicitly direct the agent to circumvent host security settings by using the -ExecutionPolicy Bypass flag when executing PowerShell scripts. This allows the execution of arbitrary, unsigned code that would otherwise be blocked by the operating system's default security posture.
[COMMAND_EXECUTION]: The provided Install-ToHarness.ps1 script and associated documentation (harness-install.md) automate the modification of configuration files for other AI agent platforms (e.g., ~/.grok/config.toml, .claude/skills, .cursor/skills). This enables lateral capability expansion into other workspace tools installed on the host.
[PROMPT_INJECTION]: The skill manifest and installation script reference a component named browser-read-bypass. The inclusion of tools explicitly documented as "bypasses" indicates intent to circumvent existing security or functional boundaries established by other software.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture is designed to ingest parameters from external, potentially untrusted project files such as README, .rdp, or environment variables to populate shell command arguments.
Ingestion points: SKILL.md (Parameters section), references/first-session-prompt.md.
Boundary markers: None identified; instructions suggest direct interpolation of data into commands.
Capability inventory: Windows shell execution, UI injection via user32, persistent credential storage, and cross-tool configuration modification.
Sanitization: No input validation, escaping, or sanitization logic is present for data ingested from the file system.

agent-workstation-drive