skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- cd3697e07eb5
- Files read
- 95
This skill collection teaches AI coding agents how to run the gws command-line tool to read, create, share, and manage data across Google Workspace services such as Gmail, Drive, Docs, Sheets, Calendar, Chat, and Meet. Its main risks are a shell-quoting instruction that can trigger local command execution from crafted spreadsheet tab names, workflows that send live emails when asked to create drafts or open Google Meet rooms to anyone with the link, and untrusted emails or documents tricking the agent into forwarding data.
What it can reach
- Shell commands
- Instructs the AI agent to run gws CLI commands, including gws generate-skills and gws auth login, in the local terminal.
- Network
- Reads, creates, modifies, shares, and deletes data across Google Workspace and Google Cloud APIs including Gmail, Drive, Docs, Sheets, Slides, Calendar, Chat, Meet, Forms, Tasks, Keep, Contacts, Classroom, Apps Script, Admin Reports, Pub/Sub, and Model Armor.
- Files
- Uploads local files to Google Drive, Gmail attachments, Google Chat, and Apps Script projects (including all non-hidden .gs, .js, .html, and appsscript.json files in a target directory), and writes downloaded files or event streams to local paths.
- Secrets
- Uses stored gws OAuth tokens or a service account key file pointed to by GOOGLE_APPLICATION_CREDENTIALS to authenticate to Google APIs.
- Browser
- Opens the system browser for interactive OAuth sign-in when running gws auth login.
Findings
-
MediumShell tip instructs agents to double-quote spreadsheet ranges, enabling command substitution on untrusted tab names
The shared reference file tells agents to wrap --range arguments in double quotes instead of single quotes to avoid zsh history expansion. In sh, bash, and zsh, single quotes already prevent expansion, whereas double quotes enable $VAR, backtick, and $(...) command substitution. Because Google Sheets allows $(...) and backticks in tab names, an agent that reads a shared spreadsheet and interpolates a crafted tab name into a double-quoted --range argument can execute arbitrary local shell commands or expose environment variables before gws runs.
Fix: Instruct agents to wrap --range and other arguments in single quotes (escaping literal single quotes as '\'') and never interpolate untrusted sheet or tab names inside double-quoted shell strings.
-
LowGoogle Meet creation recipe defaults to OPEN access, disabling lobby protection
skills/recipe-create-meet-space/SKILL.md:25
When creating a Google Meet space, this recipe hardcodes "accessType": "OPEN". In the Google Meet API, OPEN overrides organizational defaults so that anyone with the meeting link or dial-in number—including external and unauthenticated users—can join directly without knocking or host approval, even for internal team meetings.
Fix: Omit accessType so the Google Workspace organization's default policy applies, or default to TRUSTED or RESTRICTED unless the user explicitly asks for an open meeting.
-
LowEmail drafting instructions invoke gws gmail +send without --draft, sending live emails immediately
skills/persona-exec-assistant/SKILL.md:34
Both persona-exec-assistant ("Draft replies with gws gmail +send") and recipe-draft-email-from-doc ("Draft a Gmail Message from a Google Doc") instruct the agent to run gws gmail +send without the --draft flag. Without --draft, gws gmail +send transmits the email immediately instead of saving a Gmail draft, so asking the agent to draft an email or reply can send live external messages unintentionally.
Fix: Pass the --draft flag (gws gmail +send --draft ...) in all skills, personas, and recipes that describe drafting emails or replies.
-
LowShared skill encourages agents to star the GitHub repository and comment on public issues
Every gws skill directs the agent to read gws-shared/SKILL.md, which includes a section encouraging "agents/users" to star https://github.com/googleworkspace/cli and instructing agents to add context by commenting on existing public GitHub issues. In coding agents authenticated with the gh CLI or a GitHub tool, this can nudge an autonomous agent to star the repository on the user's account or post Workspace error details to a public GitHub issue without being asked.
Fix: Remove instructions aimed at agents to star the repository, and require explicit user confirmation with Workspace identifiers redacted before posting to public GitHub issues.
-
LowVacation responder recipe disables contact and domain restrictions by default
skills/recipe-create-vacation-responder/SKILL.md:24
This recipe explicitly sets both restrictToContacts and restrictToDomain to false when enabling an out-of-office auto-reply. Gmail will then send the auto-reply—typically containing travel dates and alternate internal contact emails—to every external sender, including mailing lists, spam, and reconnaissance probes, without warning the user.
Fix: Set restrictToDomain or restrictToContacts to true by default unless the user specifically asks to auto-reply to all external senders.
-
LowEmail forward and reply skills omit inline confirmation rules and do not declare gws-shared in requires.skills
skills/gws-gmail-forward/SKILL.md:16
Unlike persona and recipe skills that declare skill dependencies in metadata.openclaw.requires.skills, none of the gws-* skills list gws-shared there, so agent runners do not automatically load its confirmation and sanitization rules. While gws-gmail-send includes an inline [!CAUTION] notice requiring user confirmation before sending, gws-gmail-forward, gws-gmail-reply, and gws-gmail-reply-all omit that inline warning even though they send external emails, include original attachments by default on forwards, and accept --attach for arbitrary local files.
Fix: Add gws-shared to metadata.openclaw.requires.skills across all gws-* skills and include the inline [!CAUTION] user-confirmation warning on +forward, +reply, and +reply-all.
