skills
zenstory-ai/oh-story-claudecode/skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 41a9b6c855d7
- Files read
- 435
This toolkit for AI coding assistants helps authors plan, write, edit, generate covers, and research market rankings for Chinese web novels. Its biggest risk is that its research tools copy your main Chrome profile (including saved logins and cookies) into an unprotected debug browser and pass untrusted search links into terminal commands, which could let a malicious website access your accounts or run code on your computer.
What it can reach
- Files
- Creates, edits, and deletes novel manuscripts, outlines, and cover images, writes AI assistant settings and Git hooks into project folders, saves update caches, and copies your personal Chrome profile into ~/chrome-debug-profile.
- Network
- Connects to Chinese web-novel platforms, search engines, image-generation APIs, and GitHub for update checks, and runs local servers for Chrome remote debugging and a file-editing project dashboard.
- Shell commands
- Runs terminal commands and local scripts using Bash, PowerShell, Node.js, Python, git, curl, npx, and image utilities for writing checks, browser automation, dashboard hosting, and software updates.
- Secrets
- Reads your GPT_IMAGE_API_KEY environment variable for cover generation, copies saved login data with your Chrome profile, and can extract website cookies and login tokens from the debug browser.
- Browser
- Launches and remotely controls Google Chrome on local port 9222 using your copied login sessions to open web pages, run JavaScript, and read page content, and opens your default browser for the local dashboard.
- System
- Checks and closes running Google Chrome processes after asking permission, and installs automatic hooks that run scripts at session startup and around AI tool actions.
Findings
-
HighChrome debug port disables WebSocket Origin checks while loaded with the user's Chrome profile
skills/browser-cdp/scripts/setup-cdp-chrome.js:925
When setup-cdp-chrome.js launches Chrome for browser automation or rank scraping, it copies the user's primary Chrome profile into ~/chrome-debug-profile, runs Chrome detached in the background on 127.0.0.1:9222, and passes --remote-allow-origins=*. Normally Chrome blocks websites from connecting to its local debugging WebSocket by checking the request's Origin header; --remote-allow-origins=* turns that check off for every website. If any webpage the user or scraper visits learns or guesses a DevTools WebSocket path, or if another user or process on a shared computer connects to 127.0.0.1:9222, it can control the debug browser with no password, steal cookies from the copied Chrome profile, and open file:// URLs to read local files.
Fix: Remove --remote-allow-origins=* from chromeArgs (local CLI clients that omit the browser Origin header work without it), and use a dedicated clean debug profile instead of cloning the user's primary Chrome profile.
-
HighShell and JavaScript injection when navigating to extracted search result URLs
skills/story-setup/references/templates/agents/story-researcher.md:122
When looking up background material, the story-researcher subagent (also copied into the OpenCode and Codex agent definitions) extracts a.href URLs from Google or Bing search result pages and is instructed to place the raw URL directly inside a double-quoted shell command wrapping a single-quoted JavaScript eval call. Because Chrome's URL parser leaves characters like $, (, ), and single quotes unencoded in URL paths, a crafted link in search results can break out of the quotes or trigger shell command substitution ($(...)) on the user's machine when the agent runs the template command.
Fix: Navigate with agent-browser open using safely escaped arguments or pass the URL via standard input (--stdin) instead of interpolating untrusted URLs into a double-quoted shell eval string.
-
HighResearch subagent combines untrusted web browsing, logged-in Chrome CDP, and unrestricted shell access
skills/story-setup/references/templates/agents/story-researcher.md:7
Unlike chapter-extractor, which disables Bash and enforces a hook restricting writes to _analysis_cache, story-researcher is given unrestricted Bash and Write tools (and full shell and edit permissions in OpenCode and Codex) while visiting arbitrary third-party websites through the user's logged-in Chrome debug browser (--cdp 9222) and reading page text into its prompt. Prompt-injection text on any visited webpage can trick this subagent into running shell commands, overwriting files outside 参考资料/, or reading cookies and tokens from other logged-in tabs in the shared Chrome profile.
Fix: Remove Bash and authenticated CDP access from story-researcher in favor of stateless read-only web search/fetch tools, and add a PreToolUse path guard restricting its file writes to {project_dir}/参考资料/.
-
MediumDashboard --allow-network mode exposes unauthenticated file editing and disables Host and Origin checks
skills/story/scripts/dashboard-server.mjs:107
The local Dashboard HTTP server has no password or token authentication, and passing --allow-network causes assertLocalRequest to return immediately without checking either the Host header or the Origin header. If a user starts the dashboard with --allow-network (for example to read drafts on another device on Wi-Fi), anyone on the local network or any malicious website using DNS rebinding can read, overwrite, or delete workspace files and agent hook configurations without authentication.
Fix: Require a random per-session authentication token on all Dashboard API requests and keep Origin and Host allowlist validation active even when --allow-network is enabled.
-
MediumShell command and JavaScript injection in story-researcher CDP navigation template
skills/story-setup/references/zcode/agents/story-researcher.md:122
When story-researcher uses Chrome CDP to research background facts, it extracts links (a.href) from Google or Bing search results and instructs the agent to substitute the raw URL directly into a double-quoted Bash command (and does the same with the search query in step 3.2). Because browser URL serialization leaves characters such as $(), {}, and | unencoded in query strings and fragments, and leaves single quotes unencoded in paths and fragments, following this command template on a crafted URL from search results lets Bash execute embedded $(...) commands on the user's machine or break out of the JavaScript string in eval.
Fix: Do not interpolate untrusted URLs or queries inside double-quoted shell strings or JavaScript eval strings; pass the URL safely as a single-quoted argument to agent-browser open, or pipe scripts via a single-quoted heredoc (<<'EOF') to eval --stdin with JSON-encoded values.
-
LowBrowser setup copies saved passwords, autofill data, and unrelated site sessions into the debug profile
skills/browser-cdp/scripts/setup-cdp-chrome.js:905
To reuse login sessions, setup-cdp-chrome.js copies the user's entire Chrome Default profile directory on first run—and explicitly copies Login Data, Login Data For Account (saved passwords), and Web Data (autofill and payment data) alongside Cookies on subsequent runs—into ~/chrome-debug-profile. In addition, fs.chmodSync(debugProfile, 0o700) does not restrict directory ACLs on Windows. Duplicating saved passwords and cookies for every unrelated website goes beyond session reuse for novel platforms, exposing sensitive accounts on disk and to prompt injection if a scraped page tricks the agent into running agent-browser eval.
Fix: Do not copy the user's primary Chrome Default directory, Login Data, or Web Data; create an empty ~/chrome-debug-profile and let the user sign in only to the target novel platform inside that isolated profile.
-
LowSelf-update command runs an unpinned npx package and pulls unpinned branch HEAD
When the user asks /story to check for updates and agrees to upgrade, the skill first displays the latest GitHub release tag (vX.Y.Z), but then instructs the agent to run npx skills add zenstory-ai/oh-story-claudecode -y -g. Using -y with an unpinned npx skills package downloads and executes whatever version is currently latest on npm, and omitting @vX.Y.Z on the repository installs the mutable default-branch HEAD rather than the specific release version the user just reviewed.
Fix: Pin the CLI package version and append the verified release tag (for example, zenstory-ai/oh-story-claudecode@vX.Y.Z) instead of pulling an unpinned branch HEAD with -y.
-
LowUnauthenticated local Dashboard API can read and overwrite hidden agent hook configurations
skills/story/scripts/dashboard-server.mjs:150
While /api/tree and /api/search call shouldIgnoreDirectory to hide dot-directories such as .claude, .codex, .opencode, .zcode, and .agents, resolveWorkspacePath (used by GET, PUT, and DELETE /api/file) omits that check and allows .json, .toml, .yaml, and .md files anywhere in the workspace. Because the server listens on predictable port 127.0.0.1:43110 with no authentication token, another user or untrusted process on a shared machine can read private manuscripts and overwrite .claude/settings.local.json, .codex/hooks.json, or .agents/hooks.json to run arbitrary commands when the owner next uses their coding agent.
Fix: Enforce shouldIgnoreDirectory on all path segments inside resolveWorkspacePath and require a random bearer token generated at server startup for all /api/* endpoints.
-
LowUndocumented automatic GitHub update check on session start
skills/story-setup/references/templates/hooks/session-start.sh:182
Once /story-setup deploys Claude Code hooks, session-start.sh automatically contacts api.github.com once every 24 hours at session startup and writes ~/.claude/.story-update-cache outside the project directory unless STORY_NO_UPDATE_CHECK=1 is set. Neither story-setup/SKILL.md nor deploy-claude-code.md informs the user about this background network request or how to turn it off.
Fix: Document the automatic 24-hour update check and the STORY_NO_UPDATE_CHECK=1 environment variable in story-setup/SKILL.md, or make background update checks opt-in.
-
LowWeb-browsing research subagent combines unrestricted Bash access with untrusted page text and logged-in Chrome sessions
skills/story-setup/references/zcode/agents/story-researcher.md:8
The story-researcher subagent only needs to gather public reference material and write Markdown notes under 参考资料/, yet it is granted full Bash access while reading up to 8,000 characters of raw text (document.body.innerText) from arbitrary third-party websites returned by web search. In addition, whenever Chrome CDP port 9222 is open from an earlier browser-cdp or ranking-scan task, story-researcher automatically reuses that logged-in browser to visit third-party search results. Unlike chapter-extractor (which disables Bash when reading external texts), story-researcher is exposed to indirect prompt injection from any visited webpage, which could trick it into running shell commands or reading cookies and tokens from the attached browser.
Fix: Add Bash to disallowedTools for story-researcher and use read-only search/fetch tools—or an isolated, cookie-free browser context without general shell access—when reading untrusted external web pages.
