skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 97de40dcca2f
- Files read
- 535
This project is a collection of workflow guides for AI coding agents covering software design, code review, testing, pull requests, and autonomous coding. The biggest risk is that one skill tells the agent to run an unprotected web server on your local network that forwards a secret webhook key and to install Tailscale with administrator access from an unverified script, while other skills allow posting messages and filing tickets in connected apps without asking first.
What it can reach
- Files
- Reads and edits project files, git worktrees, agent rule and skill folders in your home directory, temporary folders, and saved workspace chat transcripts, and deletes scratch or merged worktrees and build caches.
- Network
- Connects to GitHub, Cursor cloud services and webhooks, tool download servers, and linked apps such as Slack, Linear, Notion, or Datadog, and can expose an unprotected local web server on your network and Tailscale.
- Shell commands
- Runs terminal commands for builds, tests, git, GitHub, JavaScript runtimes, formal verification, profilers, and iOS simulators, and uses administrator privileges (sudo) to install and start Tailscale.
- Secrets
- Reads a Cursor webhook secret key from a local credential file, saves a copy in a local server folder, and sends it in requests to Cursor webhook endpoints.
- Browser
- Automates local web browsers and Electron desktop apps via Playwright, Cypress, and Chrome DevTools to test interfaces and record screenshots, traces, and videos.
- System
- Launches background processes, local servers, and parallel subagents, and installs developer toolchains and Tailscale networking software.
Findings
-
MediumUnauthenticated HTTP server bound to 0.0.0.0 exposes webhook proxy to LAN and cross-site requests
skills/make-bot-ui/SKILL.md:57
When the user runs /make-bot-ui, the skill instructs the agent to start a local HTTP server bound to 0.0.0.0:<port> that holds the secret Grok Bot webhook sender key and forwards button POSTs to the webhook with that key attached. Because it binds to all network interfaces rather than just the Tailscale IP and specifies no password, token, or Origin check, anyone on the same Wi-Fi or LAN, or any malicious website the user visits in their browser via cross-site POST requests, can trigger authenticated bot webhook runs directly.
Fix: Bind the server only to the Tailscale IPv4 address from tailscale ip -4 (plus 127.0.0.1 if needed) instead of 0.0.0.0, require authentication on the UI server, and validate the Origin header or a CSRF token on all POST requests.
-
LowUnpinned remote install script piped directly to sudo sh
skills/make-bot-ui/SKILL.md:89
If Tailscale is not installed when /make-bot-ui is invoked, the skill instructs the agent to download https://tailscale.com/install.sh and pipe it directly into sudo sh without version pinning, checksum verification, or prior user confirmation. Compromised upstream script content or unexpected root package installation would execute arbitrary commands with root privileges.
Fix: Ask the user before installing system software with sudo, and prefer OS package managers or a pinned release verified against a cryptographic checksum before execution.
-
LowAutonomy rules authorize unprompted external chat and ticket actions across any MCP tool
skills/poteto-mode/SKILL.md:84
Although /poteto-mode describes itself as a style for concise replies, clean code, and verified work, its Autonomy section instructs the agent to use any MCP tool and send external team chat messages and ticket updates without asking. In addition to sending data off the machine without confirmation, prompt injection inside untrusted issues, PR comments, or logs read during a task can abuse this rule to exfiltrate workspace data through connected chat or ticket MCPs.
Fix: Require user confirmation before sending external chat messages, updating tickets, or calling write-capable external MCP tools unless the user explicitly grants permission for those actions.
-
LowRead-only investigation skill spawns write-capable subagents on untrusted external data
Users invoke /why (directly or via /teach and /recall) for read-only historical explanations, but the skill requires spawning every investigator subagent with readonly: false so MCP tools remain available. This grants subagents full file-write, shell, and write-MCP capabilities while they read untrusted external GitHub issues, PR comments, Slack messages, Sentry exceptions, and logs, and skills/why/references/investigator-prompt.md lacks both write prohibitions and prompt-injection warnings.
Fix: Add explicit instructions in skills/why/references/investigator-prompt.md forbidding file edits, shell commands, and write-capable MCP calls and warning subagents to treat external content as untrusted data, and restrict subagent tools to read-only allowlists where supported.
-
LowSession reflection auto-files issues to an external backlog tracker without user approval
The /reflect skill description states that it mines the active transcript and routes learnings into edits on existing skills, which wait for user approval. However, step 5 instructs the agent to file every synthesized Backlog item to an external team tracker immediately without waiting for approval, sending session-derived information off the machine unasked.
Fix: Present Backlog items alongside Accepted skill edits and wait for explicit user approval before filing issues to any external tracker.
