skills
pbakaus/impeccable/.agents/skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- c4cd1924c389
- Files read
- 63
Impeccable is an AI coding skill that helps agents design, review, audit, and live-edit web and mobile user interfaces. Its biggest risks are that its auto-downloaded helper launcher can be tricked into running malicious files hidden inside an untrusted project folder, and scripts on an inspected webpage can hijack its live-editing connection to read your source files or send commands to the AI agent.
What it can reach
- Files
- Reads and edits project UI code and design documents, stores downloaded helper programs in your home folder, and writes hook settings and shortcuts for AI editors like Claude, Cursor, Codex, and Gemini.
- Network
- Downloads its helper program from GitHub on first run, checks for updates, runs local servers, and sends network requests for usage tracking and image generation.
- Shell commands
- Runs its launcher scripts and helper program, starts local development servers, and runs mobile simulator tools like adb and xcrun.
- Browser
- Opens browser tabs, injects live-inspection and editing scripts into local web pages, and connects to desktop apps through local debugging ports.
Findings
-
HighWindows launcher executes untrusted impeccable binary or script from the current working directory
.agents/skills/impeccable/scripts/impeccable.cmd:105
SKILL.md instructs the agent to run impeccable.cmd with the current working directory set to the target project. On first run or immediately after an engine version bump (before the version-pinned cache is populated), impeccable.cmd runs 'where impeccable' to check PATH. On Windows, where.exe searches the current working directory before PATH unless '$PATH:' is specified. If a cloned or untrusted project contains impeccable.cmd, impeccable.bat, impeccable.exe, or impeccable.com in its root folder, the launcher immediately executes it during the :probe check without verifying any checksum.
Fix: Change 'where impeccable' to 'where $PATH:impeccable' so where.exe never searches the current working directory, and verify candidate binaries against engine-checksums.txt before running them.
-
HighScripts on the inspected page can capture the Live Mode token and send steer commands to the AI agent
.agents/skills/impeccable/scripts/live-browser.js:18
Live Mode injects live.js into the inspected page right before </body> and passes the helper server's secret token via window.__IMPECCABLE_TOKEN__. Although live-browser.js deletes that global, scripts already running on the page (from an untrusted project, a compromised third-party script, or page XSS) can intercept it with an Object.defineProperty setter before live.js runs, wrap window.fetch to read the token from the 5-second /status?token= poll, or drive the DOM-mounted Steer input directly. With the token, page JavaScript can fetch project files via /source?token= and post forged 'steer' events that live.md instructs the polling AI coding agent to execute as direct user instructions.
Fix: Isolate the Live Mode UI and token in a separate origin (such as a browser extension or a cross-origin parent frame) rather than the inspected page's JavaScript context, and require terminal confirmation before the agent acts on steer events.
-
MediumUnsanitized HTML injection in the Live Mode DESIGN.md component preview panel
.agents/skills/impeccable/scripts/live-browser.js:13497
When the user opens the DESIGN.md panel in Live Mode, renderComponentTiles() takes components[].html from .impeccable/design.json and assigns it directly to container.innerHTML inside an open shadow root without sanitization. Shadow DOM isolates CSS only, not JavaScript execution. Even on a static HTML page with no scripts of its own, a crafted .impeccable/design.json in a repository or pull request (such as an img tag with an onerror handler) executes arbitrary JavaScript in the page origin and can access app storage or capture the Live Mode helper token.
Fix: Render component snippets inside a sandboxed iframe without allow-scripts, or sanitize c.html with a strict allowlist that removes script elements, event-handler attributes, and javascript: URLs.
-
MediumSkill instructions direct the agent to bypass sandboxes for dependency installation and dev servers
.agents/skills/impeccable/reference/live.md
live.md instructs Codex to run not only the skill's helper commands, but also the target app's dev server and any dependency-installing setup (such as npm install), with sandbox_permissions set to 'require_escalated' from the start. Similarly, new-work.md tells the agent to choose the least-sandboxed command path, and hooks.md says never to call the engine sandboxed. Running third-party package install scripts and dev servers outside the agent sandbox exposes the user's host files and credentials when inspecting an untrusted repository.
Fix: Do not instruct the agent to escalate out of the sandbox for project dependency installation or dev servers by default; keep untrusted project scripts sandboxed unless the user explicitly chooses otherwise.
-
LowLauncher downloads and executes an external compiled binary at runtime
.agents/skills/impeccable/scripts/impeccable:231
When no preinstalled binary is present, the shell and Windows launchers download a compiled native executable from GitHub Releases into ~/.impeccable/bin on first use and run it (and invoke it on file edits when hooks are enabled). While the launchers enforce a pinned version and fail-closed SHA-256 verification against engine-checksums.txt, the binary's source code is not part of the skill folder and cannot be audited alongside the skill text.
Fix: Build the helper from auditable source included in the repository or publish reproducible build provenance alongside engine-checksums.txt.
-
LowUndisclosed telemetry ping sent during design direction selection
.agents/skills/impeccable/reference/new-work.md:102
During new-work design selection, the skill instructs the agent to execute a 'concept-seed --kind' choice-ping command whenever concept-seed prints a TELEMETRY line, and context also performs update checks. Outbound telemetry and how to disable it are not disclosed to the user in SKILL.md.
Fix: Make telemetry strictly opt-in, or clearly document what data leaves the machine and provide a configuration flag and environment variable to disable it.
