The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions explicitly promote downloading and executing remote shell scripts via curl | bash and irm | iex (PowerShell) from an unverified third-party GitHub repository.
Evidence: curl -fsSL https://raw.githubusercontent.com/pranshuparmar/witr/main/install.sh | bash in SKILL.md.
Evidence: irm https://raw.githubusercontent.com/pranshuparmar/witr/main/install.ps1 | iex in SKILL.md.
[PRIVILEGE_ESCALATION]: The skill suggests commands that require administrative permissions, including bypassing package integrity checks and running system-level process inspection with sudo.
Evidence: sudo apk add --allow-untrusted ./witr-*.apk in SKILL.md.
Evidence: sudo witr <pid> recommended for troubleshooting in SKILL.md.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of binaries and scripts from a personal GitHub repository (pranshuparmar/witr) that is not affiliated with the skill author or a trusted vendor list.
Evidence: Multiple installation methods pointing to github.com/pranshuparmar/witr.
[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and processing untrusted system data (process names, port metadata, service unit files) which could potentially contain malicious instructions if a system is already partially compromised.
Ingestion points: Reads process metadata, binary paths, and systemd unit files via the witr tool.
Boundary markers: None; the agent is not instructed to treat tool output as untrusted.
Capability inventory: The agent is encouraged to execute shell commands, perform package installations, and write to the filesystem.
Sanitization: No evidence of output sanitization before the agent processes the causality chain data.

witr-process-inspector