superpowers
- Type
- Plugin
- Scanned
- Nov 13, 2026
- Version
- becc25649700
- Files read
- 133
This plugin adds over 75 AI agent 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-on skills pre-approve unrestricted shell commands, Python and Node execution, file writes, and belt CLI calls, which could let prompt-injected web content run commands on your computer, upload local files to the cloud, or post to connected X/Twitter accounts without asking first.
What it can reach
- Shell commands
- Pre-approves running belt CLI commands across 70+ skills, plus unrestricted Bash commands (ai-podcast), node/npm/npx/pnpm/yarn (javascript-sdk), python and pip install inferencesh (python-sdk), and npx skills (related-skill).
- Files
- Reads and writes local files without confirmation in the ai-podcast skill, and automatically uploads local files to inference.sh cloud storage whenever a file path is passed to belt app run or belt app deploy.
- Network
- Sends prompts, search queries, media, and uploaded files to inference.sh cloud APIs and third-party model providers, and downloads CLI binaries, skills, and UI components from inference.sh, GitHub, and npm.
- Secrets
- Uses inference.sh login credentials and INFSH_API_KEY or INFERENCE_API_KEY environment variables, along with any third-party OAuth integrations connected to the user's inference.sh account.
- Browser
- Controls remote headless Playwright browser sessions via agent-browser (navigating pages, filling forms, running page JavaScript, taking screenshots, and recording video) and opens a local browser during belt login.
- Other
- Can post, like, retweet, and delete tweets, follow accounts, and send direct messages on connected X/Twitter accounts through inference.sh x/* apps.
Findings
-
HighOver-broad allowed-tools pre-approves arbitrary shell and code execution
guides/content/ai-podcast/SKILL.md:4
In Claude Code, allowed-tools in a skill's frontmatter lets the agent run those tools without asking your permission first. While most skills in this plugin limit shell access to belt commands, ai-podcast pre-approves unrestricted Bash, Read, and Write access; 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 untrusted text in a webpage, document, or issue tricks the agent via prompt injection while these skills are active or invokes them, the agent can run arbitrary commands, execute Python or Node code, install external packages, or overwrite files on your machine with no confirmation prompt.
Fix: Restrict guides/content/ai-podcast/SKILL.md to Bash(belt *), and remove wildcard runtime and package-runner permissions such as Bash(node *), Bash(npx *), Bash(python *), and Bash(npx skills *) from the SDK and utility skills so the user is prompted before code runs.
-
MediumPre-approved Bash(belt *) allows unprompted local file uploads and social account actions
Over 70 default-on skills—including web-search and agent-browser, which ingest untrusted web pages—set allowed-tools: Bash(belt *) so any belt command runs without user confirmation. Because belt app run automatically uploads local files whenever a file path (such as ~/.ssh/id_rsa or .env) is supplied in its JSON input, can invoke any third-party user/app on the public store, can deploy local directories via belt app deploy, and can post tweets or send direct messages via x/post-tweet and x/dm-send, a prompt-injected webpage could trick the agent into exfiltrating sensitive local files or posting from your connected X/Twitter account.
Fix: Remove broad Bash(belt *) pre-approval from skills that handle untrusted web content, or narrow allowed-tools to specific safe subcommands and app namespaces so file uploads, deployments, and social media actions require user confirmation.
-
LowPlaintext HTTP curl | sh command and unpinned remote CLI installer scripts
More than 70 skills and the README link to cli-install.md to install the belt CLI, where the primary command omits https:// (curl -fsSL cli.inference.sh | sh). Without an explicit https:// scheme, curl makes its initial request over unencrypted HTTP before following any redirect, allowing an attacker on the same Wi-Fi or local network to intercept the connection and feed malicious shell commands directly into sh. In addition, tools/infsh-cli/SKILL.md, tools/agent-tools/SKILL.md, and sdk/building-apps/SKILL.md pipe unpinned remote scripts from https://cli.inference.sh directly into sh without pinning a version or verifying the installer script hash beforehand.
Fix: Use an explicit https:// URL in cli-install.md, pin CLI downloads to specific release versions with locally verified SHA-256 hashes before execution, or recommend a package manager instead of piping curl to sh.
-
LowSkills instruct the agent to install unpinned external skills and UI code at runtime
Over 75 SKILL.md files start with a top-level callout instructing the agent to run npx skills add belt-sh/cli, even when the user only asked to use an already-installed skill. Similarly, the UI skills instruct running npx shadcn@latest add against unpinned ui.inference.sh URLs, and remotion-render recommends npx skills add remotion-dev/skills@remotion-best-practices without version or commit pinning. Combined with skills that pre-approve npx, the agent may pull and install unpinned remote instructions or code at runtime, creating a supply-chain risk if a referenced repository or endpoint is compromised.
Fix: Remove automatic npx skills add directives from the top of SKILL.md prompt bodies and pin any recommended external skills or component URLs to immutable commit hashes or release versions.
-
LowSDK tool-builder reference templates use eval() on model-generated 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 handling tool calls by passing LLM-generated arguments directly to eval() (eval(call.args.expression)). Although the installed plugin does not execute these examples itself, coding agents use these reference files as templates when writing application code, which can introduce remote code execution via prompt injection into apps built with these skills.
Fix: Replace eval() in both the JavaScript and Python tool-builder reference files with a safe math parser or explicit arithmetic operations.
-
LowUI and SDK quick-start templates expose an unauthenticated API proxy route
The setup instructions in ui/agent-ui/SKILL.md (and the default Next.js and Express proxy snippets in sdk/javascript-sdk/SKILL.md) export a public /api/inference/proxy route backed by INFERENCE_API_KEY without any user authentication or rate limiting. While the plugin itself does not start a server, deploying an app scaffolded from this template allows anyone who can reach the site to run inference.sh tasks and spend your API credits.
Fix: Include authentication and rate-limiting checks directly in the default proxy route examples in ui/agent-ui/SKILL.md and sdk/javascript-sdk/SKILL.md, and add a clear warning above unauthenticated snippets.
-
LowUnescaped JSON string concatenation and unquoted output in agent-browser shell templates
tools/utilities/agent-browser/templates/capture-workflow.sh:55
The shell templates in tools/utilities/agent-browser/templates/ (capture-workflow.sh, authenticated-session.sh, and form-automation.sh) splice URLs and credentials directly into JSON strings without escaping quotes, and pass unquoted $RESULT output from scraped pages to echo. If a user or agent runs these templates on a URL containing double quotes, extra JSON fields such as proxy_url can be injected into agent-browser, and wildcard characters in scraped webpage text trigger bash filename expansion against the current directory.
Fix: Build JSON arguments safely with jq -n --arg url "$TARGET_URL" instead of string concatenation, and double-quote "$RESULT" and "$SESSION_ID" throughout the template scripts.
