← MCP server and skill grades

superpowers

101-skills/superpowers

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

This plugin bundles over 80 skills and guides that let an AI coding agent generate media, search the web, automate browsers and X/Twitter, and build apps using the inference.sh (belt) CLI and SDKs. Its biggest risk is that several skills pre-approve broad shell, Node.js, Python, file-write, and belt CLI execution without user confirmation, allowing prompt injection in web pages or untrusted files to run commands or upload local files silently.

What it can reach

Shell commands
Pre-approves running belt CLI commands across dozens of skills, Node.js and package manager commands (node, npx, npm, pnpm, yarn) in javascript-sdk, python and pip in python-sdk, and unrestricted shell commands in ai-podcast.
Files
Pre-approves reading and writing arbitrary local files in ai-podcast, and the belt CLI and SDKs automatically upload local files whenever a file path is supplied in an app input.
Network
Sends prompts, search queries, local files, and app deployments to inference.sh and third-party model providers, and downloads CLI binaries, skills, and UI components from remote URLs.
Secrets
Uses local belt login credentials or INFSH_API_KEY and INFERENCE_API_KEY environment variables, as well as any OAuth or API secrets linked to the user's inference.sh account.
Browser
Controls remote headless Playwright sessions via agent-browser (including form filling, JavaScript execution, file uploads, and video recording) and opens a local browser during belt login.
Other
Can post, delete, like, retweet, follow users, and send direct messages on a connected X/Twitter account through x/* apps.

Findings

1 high5 low
  1. HighUnrestricted shell, file-write, and interpreter execution pre-approved in skill manifests

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

    In Claude Code, the allowed-tools frontmatter field lets an active skill run matching tools without asking the user for confirmation. The ai-podcast guide grants bare Bash and Write access with no command restrictions at all, while sdk/javascript-sdk/SKILL.md pre-approves Bash(node *), Bash(npx *), Bash(npm *), Bash(pnpm *), and Bash(yarn *), sdk/python-sdk/SKILL.md pre-approves Bash(python *), and tools/utilities/related-skill/SKILL.md pre-approves Bash(npx skills *). If the agent reads prompt-injected text while any of these default-enabled skills is active, it can execute arbitrary code, install attacker-controlled packages or skills, or overwrite local files without user approval.

    Fix: Restrict allowed-tools in guides/content/ai-podcast/SKILL.md to specific belt app commands, and remove wildcard interpreter permissions such as Bash(node *), Bash(npx *), Bash(python *), and Bash(npx skills *) so code execution and package installation require user confirmation.

  2. LowBroad Bash(belt *) auto-approval on web-facing skills allows silent local file uploads and account actions

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

    Nearly 40 skills—including web-search and agent-browser, which fetch untrusted web pages directly into the agent's context—pre-approve Bash(belt *) instead of scoping to the specific apps they need. Because the belt CLI automatically uploads any local file path provided in JSON input fields and can invoke any app or subcommand, prompt injection inside a searched or browsed page can trick the agent into uploading sensitive local files to a remote app, posting or messaging from a linked X/Twitter account (x/post-tweet, x/dm-send), or running belt app deploy without prompting the user.

    Fix: Scope allowed-tools in each skill to only the intended app prefixes (for example, Bash(belt app run tavily/*) and Bash(belt app run exa/*) in web-search) rather than allowing all belt subcommands and apps.

  3. LowCLI installation guide pipes a plaintext HTTP URL directly into sh

    cli-install.md:4

    Over 60 SKILL.md files link to cli-install.md for setting up the belt CLI, where the primary command omits the https:// scheme (curl -fsSL cli.inference.sh | sh). Without a scheme, curl makes its initial request over unencrypted HTTP on port 80 before following any redirect, allowing an attacker on the local network to intercept the request and feed arbitrary shell commands straight into sh. Other skills (such as tools/infsh-cli/SKILL.md and sdk/building-apps/SKILL.md) use https:// but still pipe unpinned remote installer scripts directly to sh.

    Fix: Always specify https:// with --proto '=https' --tlsv1.2 in curl commands, and prefer package-manager or checksum-verified release downloads over piping remote scripts directly into sh.

  4. LowSDK tool-builder references teach the agent to pass LLM tool arguments to eval()

    sdk/javascript-sdk/references/tool-builder.md:240

    Both sdk/javascript-sdk/references/tool-builder.md and sdk/python-sdk/references/tool-builder.md implement their example client-side tool handlers by calling eval() directly on model-supplied tool arguments (call.args.expression) without sandboxing or approval gates. When a user asks their coding agent to scaffold an app using these SDK skills, copying this pattern introduces remote code execution via prompt injection into the user's application.

    Fix: Replace eval() in both JavaScript and Python tool-builder reference examples with a safe arithmetic parser or restricted lookup, and warn against passing LLM tool arguments to eval().

  5. LowAgent UI Quick Start scaffolds an unauthenticated API proxy exposing INFERENCE_API_KEY

    ui/agent-ui/SKILL.md:31

    Both ui/agent-ui/SKILL.md and the proxy Quick Start in sdk/javascript-sdk/SKILL.md instruct the coding agent to export a Next.js /api/inference/proxy route backed by INFERENCE_API_KEY with no authentication, authorization, or rate limiting. Any application deployed from this template lets unauthenticated internet visitors call inference.sh models and workspace agents using the owner's API key and billing quota.

    Fix: Include session authentication and rate-limit checks directly in the default proxy route examples in ui/agent-ui/SKILL.md and sdk/javascript-sdk/SKILL.md rather than relegating authentication to a separate reference file.

  6. LowUnescaped URL and credential interpolation in agent-browser shell templates enables JSON injection

    tools/utilities/agent-browser/templates/capture-workflow.sh:55

    The ready-to-use shell scripts in tools/utilities/agent-browser/templates/ (capture-workflow.sh, authenticated-session.sh, and form-automation.sh) concatenate URLs and credentials directly into raw JSON strings and pass $RESULT unquoted to echo. A URL containing double quotes can inject extra JSON parameters such as proxy_url into the remote browser session, and unquoted echo $RESULT expands glob characters from untrusted page text against local filenames.

    Fix: Construct JSON inputs safely with jq -n --arg url "$TARGET_URL" '{url: $url}' and quote "$RESULT" (or pipe printf '%s' "$RESULT") when passing output to jq.