← MCP server and skill grades

superpowers

qu-skills/superpowers

C
Type
Plugin
Scanned
Nov 13, 2026
Version
becc25649700
Files read
133

This Claude Code plugin provides 82 skills and guides for generating images, video, audio, web searches, browser automation, and social media posts through the inference.sh (belt) CLI and SDKs. Its biggest risks are skill definitions that pre-approve unrestricted shell commands, code interpreters, file writes, and belt file-uploading or Twitter actions without confirmation, alongside a CLI install guide that pipes an unencrypted HTTP URL directly into sh.

What it can reach

Shell commands
Runs belt CLI commands, package managers (npm, npx, pnpm, yarn, pip), and local python, node, or bash commands instructed by the skills.
Files
Reads and writes local project files and media outputs, and automatically uploads local files referenced in belt app inputs to inference.sh cloud storage.
Network
Connects to inference.sh cloud APIs and third-party AI providers, fetches UI component code from ui.inference.sh, and downloads packages and skills from npm, PyPI, and GitHub.
Secrets
Uses local belt login credentials and environment variables such as INFSH_API_KEY and INFERENCE_API_KEY, plus any OAuth integrations configured in inference.sh.
Browser
Controls a remote headless Playwright browser (agent-browser) capable of visiting URLs, filling login forms, executing page JavaScript, and recording sessions.
Other
Posts tweets, uploads media, likes, retweets, follows accounts, deletes posts, and sends direct messages on a connected X/Twitter account via x/* apps.

Findings

1 high1 medium3 low
  1. HighSkills pre-approve unrestricted shell, interpreter, and file-write tools

    guides/content/ai-podcast/SKILL.md:4

    In Claude Code, the allowed-tools frontmatter pre-authorizes listed tools to run without asking the user for confirmation while that skill is active. The ai-podcast guide grants unrestricted Bash and Write access, while sdk/python-sdk/SKILL.md grants Bash(python *), sdk/javascript-sdk/SKILL.md grants Bash(node *), Bash(npx *), Bash(npm *), Bash(pnpm *), and Bash(yarn *), and tools/utilities/related-skill/SKILL.md grants Bash(npx skills *). If the agent reads untrusted text containing a prompt injection while any of these default-on skills is active, it can execute arbitrary commands, install attacker packages, or overwrite files on the user's computer without a permission prompt.

    Fix: Restrict allowed-tools in guides/content/ai-podcast/SKILL.md to Bash(belt *) to match the other media guides, and remove wildcard interpreter and package-runner entries such as Bash(python *), Bash(node *), and Bash(npx *) so the user is prompted before local code execution.

  2. MediumCLI install command uses plaintext HTTP piped directly to sh

    cli-install.md:4

    README.md and more than 60 skill files direct users and agents to cli-install.md to install the required belt CLI, where the primary command omits the https:// scheme (curl -fsSL cli.inference.sh | sh). Without an explicit scheme, curl makes its initial request over unencrypted HTTP, allowing an attacker on the same Wi-Fi or local network to intercept the request and return malicious shell commands that sh executes immediately, bypassing the binary checksum check described below it. Several other skill files also pipe unpinned remote install scripts (such as https://cli.inference.sh) directly into sh or bash without verifying a pinned version or script hash first.

    Fix: Use an explicit https:// URL with --proto '=https' --tlsv1.2, pin release versions, and download and verify archive checksums before executing any installer.

  3. LowWildcard Bash(belt *) permission in web-reading skills allows unprompted file uploads and account actions

    tools/llm/web-search/SKILL.md:4

    More than 70 skills—including web-search, ai-rag-pipeline, and agent-browser, which are designed to fetch and process untrusted webpages—pre-approve Bash(belt *) rather than scoping permissions to specific search or browser apps. Because belt automatically uploads any local file path passed in JSON inputs (such as ~/.ssh keys or .env files) to inference.sh cloud storage, deploys local directories via belt app deploy, and performs public or destructive X/Twitter actions (x/post-tweet, x/dm-send, x/post-delete), a prompt injection on a visited webpage can trigger silent data exfiltration or unauthorized social media posts without user confirmation.

    Fix: Narrow allowed-tools in each skill to only the subcommands and app namespaces it needs (for example, Bash(belt app run tavily/*) and Bash(belt app run exa/*) in web-search), and require user confirmation for publishing, deployment, or file-uploading commands.

  4. LowSDK tool-builder templates pass LLM tool arguments directly to eval()

    sdk/python-sdk/references/tool-builder.md:365

    Both sdk/python-sdk/references/tool-builder.md and sdk/javascript-sdk/references/tool-builder.md provide client-side tool handler templates that pass untrusted LLM tool-call arguments directly into Python and JavaScript eval(), even in complete agent examples that also enable web search and omit human approval on the calculator tool. Although the plugin itself does not run this reference code on install, an application scaffolded by the coding agent from these templates would allow prompt injection in user input or search results to execute arbitrary code on the host running the client.

    Fix: Replace eval() in both Python and JavaScript tool-builder reference files with a safe arithmetic parser so LLM-generated arguments are never evaluated as code.

  5. LowDefault frontend proxy templates expose INFERENCE_API_KEY without authentication

    ui/agent-ui/SKILL.md:31

    The setup instructions in ui/agent-ui/SKILL.md and sdk/javascript-sdk/SKILL.md tell the coding agent to create a /api/inference/proxy route backed by INFERENCE_API_KEY with no authentication, rate limiting, or app allowlist by default. While this example code does not run when the plugin is installed, deploying a web app scaffolded from this setup lets any unauthenticated internet visitor call arbitrary inference.sh models and consume the owner's API credits through the proxy endpoint.

    Fix: Include session authentication, rate limiting, and allowed app checks directly in the default proxy route examples in ui/agent-ui/SKILL.md and sdk/javascript-sdk/SKILL.md.