superpowers
- Type
- Plugin
- Scanned
- Nov 13, 2026
- Version
- becc25649700
- Files read
- 133
This Claude Code plugin bundles over 80 skills and guides for generating images, video, audio, web searches, browser automation, and social media posts through the inference.sh cloud CLI and SDKs. Its biggest risk is that several default skills pre-approve unrestricted shell, interpreter, and CLI commands without user confirmation—letting a prompt-injected agent run arbitrary code, upload local files, or act on linked accounts—while its CLI install guide pipes a plaintext HTTP script directly to sh.
What it can reach
- Shell commands
- Skills instruct the agent and pre-approve running the belt CLI, npm, npx, node, pnpm, yarn, pip, python, and unrestricted bash commands.
- Files
- Skills can read and write local files, upload any local file path passed in belt or agent-browser inputs to inference.sh, and upload project directories during app deployment.
- Network
- Sends prompts, files, and URLs to inference.sh cloud workers and model providers (such as OpenRouter, Google, OpenAI, ElevenLabs, Tavily, Exa, and X), and downloads CLI binaries and packages from inference.sh, npm, and PyPI.
- Secrets
- Uses stored belt login credentials, INFSH_API_KEY and INFERENCE_API_KEY environment variables, connected OAuth integrations, and user-supplied login or proxy passwords.
- Browser
- Opens the local browser during belt login and controls a remote headless Playwright browser on inference.sh to visit pages, fill forms, execute JavaScript, and record screenshots or video.
- Other
- Can post, like, retweet, and delete tweets, follow accounts, and send direct messages on the user's connected X/Twitter account via inference.sh apps.
Findings
-
HighOver-broad allowed-tools settings auto-approve arbitrary shell and code execution
guides/content/ai-podcast/SKILL.md:4
In Claude Code, tools listed in a skill's allowed-tools frontmatter run without asking the user for permission. While most skills in this plugin restrict commands to Bash(belt *), the ai-podcast guide grants completely unrestricted Bash and Write access. Similarly, sdk/javascript-sdk/SKILL.md auto-approves Bash(node *), Bash(npx *), and Bash(npm *), sdk/python-sdk/SKILL.md auto-approves Bash(python *), and tools/utilities/related-skill/SKILL.md auto-approves Bash(npx skills *). If untrusted text tricks the agent via prompt injection while any of these default-enabled skills is active, the agent can run arbitrary local code or overwrite files with no confirmation prompt.
Fix: Change allowed-tools in guides/content/ai-podcast/SKILL.md to Bash(belt *) and remove wildcard interpreter permissions such as Bash(node *), Bash(npx *), and Bash(python *) from the SDK and utility skills so local code execution requires user approval.
-
MediumBrowser automation guide sends plaintext passwords and 2FA codes via CLI arguments to remote workers
tools/utilities/agent-browser/references/authentication.md:37
The agent-browser skill triggers on generic terms like playwright, headless browser, and fill form, but runs on remote inference.sh cloud workers rather than locally. Its authentication reference and shell template instruct passing user passwords, Google SSO passwords, proxy credentials, and 2FA codes directly inside the --input command-line argument of belt app run. This exposes secrets in local process listings (ps), transmits plaintext credentials and screenshots to remote inference.sh task records without warning, and breaks JSON syntax if a password contains quotes or backslashes.
Fix: Warn clearly that agent-browser executes on remote inference.sh servers, avoid sending third-party SSO passwords or 2FA codes in plaintext task inputs, and construct JSON safely with jq into a restricted temporary file instead of CLI argument interpolation.
-
LowCLI install guide pipes plaintext HTTP and unpinned remote scripts into sh
The CLI install guide linked across more than 60 skills runs curl -fsSL cli.inference.sh | sh without an https:// scheme. Without a scheme, curl makes its initial connection over unencrypted HTTP on port 80, allowing an attacker on the local network (such as public Wi-Fi) to intercept the request before any HTTPS redirect and execute arbitrary shell commands. Even where https:// is specified in other skill files or npx commands like npx skills add belt-sh/cli and npx shadcn@latest add are used, the remote scripts and packages are unpinned.
Fix: Always include https:// (and --proto '=https' --tlsv1.2) on curl URLs, and download version-pinned release archives verified against pinned SHA-256 checksums instead of piping remote scripts to sh or running unpinned @latest packages.
-
LowWildcard Bash(belt *) permission on search and media skills permits unprompted file uploads and account actions
tools/llm/web-search/SKILL.md:4
Over 60 skills—including web-search, ai-rag-pipeline, and agent-browser, which routinely ingest untrusted web content—pre-approve Bash(belt *) rather than scoping to the specific app they need. Because the belt CLI automatically uploads any local file path supplied in JSON inputs, posts or deletes content on linked X/Twitter accounts (x/post-tweet, x/dm-send), deploys local directories (belt app deploy), and executes local app code (belt app test), a prompt injection in a search result can exfiltrate local files or act on connected accounts without confirmation.
Fix: Scope allowed-tools in each skill to only the specific subcommands and app namespaces that skill needs, such as Bash(belt app run tavily/*) and Bash(belt app run exa/*) for web-search.
-
LowSDK tool-builder references instruct the agent to use eval() on LLM tool arguments
sdk/javascript-sdk/references/tool-builder.md:240
Both sdk/javascript-sdk/references/tool-builder.md and sdk/python-sdk/references/tool-builder.md demonstrate client-side tool handlers (which run on the developer's own machine or server) by calling eval() directly on call.args.expression from the model with no approval gate. Although the installed plugin does not run this example itself, a coding agent following this skill when building a user's app can introduce arbitrary code execution triggered via prompt injection.
Fix: Replace eval() in both the JavaScript and Python tool-builder reference examples with a safe arithmetic parser or restricted library.
-
LowUI and SDK setup guides scaffold an unauthenticated API proxy route exposing INFERENCE_API_KEY
The setup steps in ui/agent-ui/SKILL.md and sdk/javascript-sdk/SKILL.md instruct the coding agent to add an /api/inference/proxy route backed by process.env.INFERENCE_API_KEY with no authentication, rate limiting, or app allowlist (authentication is only shown as an optional pattern in a separate server-proxy.md reference). While the plugin itself does not run a server, apps deployed from this scaffold let any anonymous network caller run inference.sh tasks billed to the owner's API key.
Fix: Include authentication and authorization checks directly in the default proxy route examples in ui/agent-ui/SKILL.md and sdk/javascript-sdk/SKILL.md, and warn against deploying an open proxy.
