skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 3d82c85419ad
- Files read
- 522
This collection of skills lets AI coding agents read and manage your Feishu or Lark workspace, including chats, emails, documents, spreadsheets, calendars, approvals, and Miaoda web apps, using the lark-cli tool. The biggest risk is that some workflows skip confirmation prompts and silently download and run unpinned software from npm, while untrusted content in chats, documents, or shared apps can trick the agent into uploading local or clipboard files, running local scripts, or changing permissions and approving requests on your behalf.
What it can reach
- Shell commands
- Runs lark-cli commands, Git and npm commands, curl, jq, local Python and Node.js helper scripts, and npx to download and run the whiteboard CLI tool.
- Files
- Reads and writes local files inside and outside the working folder to upload or download attachments, sync Drive folders, export data, clone Miaoda app code, and update .env.local files.
- Network
- Connects to Feishu, Lark, and Miaoda APIs and live event streams, fetches online API documentation and remote media URLs, and downloads packages from npm.
- Secrets
- Uses and manages Feishu and Lark login tokens and bot credentials, local Git credentials, Miaoda app environment variables, OpenAPI keys, and webhook tokens.
- System
- Reads images directly from your computer's clipboard when inserting media or covers into documents.
- Other
- Reads, creates, edits, shares, and deletes data across your Feishu or Lark account, including messages, emails, cloud files and permissions, calendars, meetings, wikis, tasks, corporate approvals, and Miaoda apps and databases.
Findings
-
MediumWhiteboard skill silently downloads and runs an unpinned npm package
skills/lark-whiteboard/SKILL.md:15
At the start of any whiteboard task, the skill instructs the AI agent to run npx -y @larksuite/whiteboard-cli@^0.2.13 -v without asking the user. Passing -y auto-installs packages without confirmation, and the caret range ^0.2.13 is unpinned, meaning any newly published 0.2.x release runs automatically. In addition, npx reads .npmrc from the current working directory, so running the agent inside a cloned untrusted repository with a malicious .npmrc can redirect npx to an attacker-controlled registry and execute arbitrary code. This runtime download is also omitted from metadata.requires.bins.
Fix: Pin @larksuite/whiteboard-cli to an exact version, declare whiteboard-cli in metadata.requires.bins so users install it explicitly ahead of time, and remove npx -y auto-installation without user consent.
-
MediumBackground card callback listener feeds untrusted chat form input directly into the agent
skills/lark-im/references/lark-im-card-action-reply.md:101
When starting a card.action.trigger listener, this reference instructs the agent to offer automatic background handling of incoming card events. Once enabled, free-form text submitted by any chat participant in input_value, form_value, or action_value is streamed straight into the AI coding agent for immediate processing, with no prompt-injection warning and no rule restricting the agent to calling only the card update API. A stranger in a group chat can submit prompt-injection text in a card form field to trick the local agent into running shell commands or accessing the user's files and Feishu data.
Fix: Explicitly treat operator-supplied fields (input_value, form_value, option, options, action_value) as untrusted data, forbid executing instructions found inside them, and restrict the auto-update loop strictly to constructing card JSON and calling /open-apis/interactive/v1/card/update.
-
MediumRelative scripts/ invocations can execute untrusted Python files from the current working directory
skills/lark-sheets/references/lark-sheets-read-data.md:59
Multiple spreadsheet and slide instructions (including lark-sheets-read-data.md, lark-sheets-chart.md, lark-sheets-workbook.md, and lark-slides/SKILL.md) tell the agent to check for or run helper scripts using relative paths such as python3 scripts/lark_inspect_workbook.py and python3 scripts/lark_chart_quality_check.py instead of resolving the installed skill's own directory. Because the agent's working directory is typically the project folder the user opened, running a read-only Feishu sheet inspection or chart check while inside an untrusted cloned repository or download folder that contains matching scripts/lark_*.py files will execute attacker-controlled Python code on the user's machine.
Fix: Reference helper scripts using the absolute path of the installed skill directory rather than relative scripts/... paths in the current working directory, and never add relative "scripts" to sys.path.
-
MediumEditing reply or forward drafts with set_body auto-embeds local files referenced in quoted HTML
skills/lark-mail/references/lark-mail-draft-edit.md:25
The draft editor automatically resolves any relative <img src="./..."> tag without a URI scheme in set_body (and in the --body flag, which expands to set_body) by reading that relative path from the current working directory and embedding the local file as an inline MIME part. At the same time, this table instructs the agent to pass the full HTML including the quoted original email into set_body when editing quoted content on a reply or forward draft. If an incoming HTML email from a stranger contains a relative <img src="./..."> pointing to a file in the working directory, re-submitting the quoted HTML via set_body silently attaches that local file into the draft.
Fix: Warn the agent to strip or neutralize any relative <img src> attributes inside untrusted quoted HTML before passing it to set_body or --body, or restrict local image embedding strictly to newly authored content outside quoted blocks.
-
MediumAmbiguous media insertion tries --from-clipboard first and uploads unseen clipboard images
skills/lark-doc/references/lark-doc-media-insert.md:34
When the user does not clearly specify whether an image comes from a file or the clipboard, this fallback rule instructs the agent to try --from-clipboard first, streaming the OS clipboard image directly from memory into the cloud document without saving or previewing it. If the user previously copied a sensitive screenshot (such as credentials or private messages), or if a prompt injection triggers docs +media-insert, the clipboard image is immediately uploaded to a potentially shared Feishu document.
Fix: Require explicit user confirmation before reading from --from-clipboard when the image source was not explicitly stated as the clipboard by the user.
-
MediumWhiteboard skill downloads and runs an unpinned npm package at runtime
skills/lark-whiteboard/references/lark-whiteboard-update.md:92
The whiteboard skill (across lark-whiteboard-update.md, lark-whiteboard-workflow.md, routes/*.md, and elements/*.md) instructs the AI agent to run npx -y @larksuite/whiteboard-cli@^0.2.13 to render previews, check SVGs, list icons, and convert diagrams. Passing -y skips confirmation to download and execute code from the public npm registry at runtime, while the caret range (^0.2.13) automatically pulls any newly published 0.2.x version and its transitive dependencies without declaring this binary in the skill metadata.
Fix: Declare whiteboard-cli in metadata.requires.bins so users pre-install a vetted version, or remove -y and pin @larksuite/whiteboard-cli to an exact version.
-
MediumUnpinned npm package downloaded and executed at runtime without confirmation
skills/lark-whiteboard/scenes/mention.md:38
The whiteboard scene guides (mention.md, line-chart.md, and treemap.md) instruct the agent to run npx -y @larksuite/whiteboard-cli@^0.2.13 in the shell. Passing -y skips confirmation prompts while the caret range (^0.2.13) allows any newly published 0.2.x release or updated transitive dependency from npm to be downloaded and executed on the user's machine at runtime.
Fix: Declare whiteboard-cli as a required pre-installed binary in metadata.requires.bins, or pin an exact version without a caret range and remove -y so the user must approve package installation.
-
MediumSlides error-recovery recipe silently switches to user identity and grants bot full_access with --yes
skills/lark-slides/references/cli/lark-slides-media-upload.md:90
When a command running under the restricted --as bot identity fails with 1061004 (forbidden), the troubleshooting table suggests switching to --as user and running drive +member-add with --perm full_access and --yes to grant the bot management access to the presentation. If an agent follows this recovery step on an untrusted presentation token, it crosses the bot-to-user privilege boundary and bypasses drive +member-add's --yes confirmation gate without asking the user.
Fix: Require explicit user confirmation before switching from --as bot to --as user or passing --yes to drive +member-add.
-
MediumWiki space member grant lacks confirmation gate and allows silent admin addition
skills/lark-wiki/references/lark-wiki-member-add.md:5
Even though Feishu's underlying members.create API is marked danger: true, wiki +member-add (and +member-remove) omits the --yes confirmation gate used on drive +member-add and apps +member-add, omits the lark-shared safety prerequisite, and supports --need-notification=false to suppress notifications. Treating an admin or member grant on a wiki space or personal library (my_library) as harmlessly reversible ignores that an attacker added via prompt injection can immediately read or copy all confidential pages before removal.
Fix: Require --yes confirmation on wiki +member-add and +member-remove, link lark-shared/SKILL.md, and instruct the agent to confirm the target space, member, role, and notification setting with the user before execution.
-
LowDocument and meeting skills instruct the agent to skip shared safety rules
The project centralizes its safety rules in lark-shared/SKILL.md, including confirming user intent before write or delete actions, previewing risky calls with --dry-run, respecting exit-code-10 high-risk approval gates, and never printing secrets. However, lark-doc/SKILL.md (and similarly lark-meeting/SKILL.md) explicitly tells the agent not to read lark-shared/SKILL.md unless an authentication or scope error occurs. For an already-logged-in user, the agent handles untrusted cloud documents and meeting chats without loading those safety guardrails, while lark-doc also permits @/absolute/path references to upload local files outside the working directory.
Fix: Require reading lark-shared/SKILL.md before running write or local-file upload commands, or inline the mandatory write-confirmation and path-safety rules directly in lark-doc/SKILL.md and lark-meeting/SKILL.md.
-
LowBase field extension guide pre-authorizes high-risk bulk writes with --yes
skills/lark-base/references/lark-base-field-extension.md:34
lark-cli protects high-risk write commands with an exit-code-10 confirmation gate so the agent pauses and asks the user before passing --yes. However, lark-base/SKILL.md does not load lark-shared/SKILL.md, and lark-base-field-extension.md states that +field-extension-update and +field-extension-update-cells must always be called with --yes (including --yes in every command example, such as overwriting an entire column via --type column) without instructing the agent to obtain explicit user confirmation first. If the agent misinterprets a request or is tricked via prompt injection in Base records, the CLI confirmation gate is bypassed on the initial call.
Fix: Remove --yes from initial command examples and instruct the agent to confirm the target table, field, and row or column scope with the user before passing --yes.
-
LowStarter deck component posts slide HTML and speaker notes to wildcard origin
skills/lark-apps/creative-design/starter-components/deck-stage.js:771
The deck-stage.js starter component copied into Miaoda HTML slide decks sends full speaker notes (miaoda:deck:notes-changed) and the entire deck markup this.outerHTML (miaoda:deck:deck-changed) to window.parent using a wildcard targetOrigin ('*'), and processes incoming window message events (_onMessage) without verifying event.origin. If a published or locally hosted deck is embedded in an iframe on an untrusted page without frame-ancestors restrictions and the user interacts with it, the framing page can send control messages and read the slide HTML and speaker notes cross-origin.
Fix: Restrict postMessage targetOrigin to the expected Miaoda host origin and validate event.origin and event.source === window.parent in _onMessage.
-
LowBatch spreadsheet reference instructs always passing --yes, bypassing the exit-10 confirmation gate
skills/lark-sheets/references/lark-sheets-batch-update.md:10
lark-shared and lark-sheets/SKILL.md define exit code 10 on high-risk-write commands as a mandatory confirmation gate where --yes may only be added after --dry-run preview and explicit user consent. However, lark-sheets-batch-update.md states as a hard tool constraint that --yes must always be included on +batch-update to avoid exit code 10, even though +batch-update can bundle destructive actions like +sheet-delete, +dim-delete, and +cells-clear.
Fix: Replace the '--yes 必带' instruction with the standard high-risk approval flow: run --dry-run first, show the planned batch operations to the user, and pass --yes only after explicit user confirmation.
-
LowPredictable /tmp/tpl-patch.json path used for mail template updates
skills/lark-mail/references/lark-mail-template-update.md:27
The command examples write a patch skeleton to a fixed /tmp/tpl-patch.json path in world-writable /tmp and later apply it via --patch-file /tmp/tpl-patch.json. On shared multi-user machines, a local user could pre-create or tamper with /tmp/tpl-patch.json (for example, injecting bccs recipients or altering template_content) before the update runs, and the absolute /tmp path conflicts with lark-shared's CWD-relative path rule.
Fix: Write temporary patch files inside the current working directory (such as ./tpl-patch.json) or a private mktemp file rather than a fixed path in /tmp.
-
LowPublished HTML app template loads external CDN scripts without Subresource Integrity
skills/lark-apps/creative-design/assets/index.html:9
The starter HTML template published to Miaoda web apps loads React, ReactDOM (using unminified development builds), and @babel/standalone from sf3-scmcdn-cn.feishucdn.com with crossorigin="anonymous" but without integrity hashes (and charts.md / animated-video.md do the same for ECharts and Popmotion). Without Subresource Integrity, a compromise of the CDN path would execute arbitrary JavaScript in viewers' browsers on the published app origin.
Fix: Switch to production builds of React and ReactDOM and add sha384 integrity attributes to all external CDN script tags.
