superpowers
- Type
- Plugin
- Scanned
- Nov 13, 2026
- Version
- becc25649700
- Files read
- 133
This Claude Code plugin is a fork of the inference.sh skills collection (plus an agent self-regulation guide) that teaches coding agents to generate images, videos, and audio, search the web, automate browsers and X/Twitter, and build apps via the belt CLI and SDKs. Its biggest risk is that several default-on skills pre-approve unrestricted shell, Node.js, Python, npx, and belt CLI commands without user confirmation, which could let a prompt-injected agent run arbitrary code or upload local files.
What it can reach
- Shell commands
- Skills pre-authorize running shell commands without confirmation via allowed-tools, including unrestricted Bash in ai-podcast, node/npx/npm/pnpm/yarn in javascript-sdk, python/pip in python-sdk, npx skills in related-skill, and belt CLI commands across nearly 70 skills.
- Files
- The ai-podcast skill pre-authorizes reading and writing local files, and the belt CLI and SDKs automatically upload local files to inference.sh when given a local file path.
- Network
- Sends prompts, media, and queries to inference.sh and upstream AI/search providers (such as OpenRouter, Google, ElevenLabs, Tavily, Exa, fal.ai, ByteDance, and Alibaba), and fetches CLI binaries, UI components, and packages from inference.sh domains, GitHub, npm, and PyPI.
- Secrets
- Uses stored belt CLI login tokens, INFSH_API_KEY or INFERENCE_API_KEY environment variables, and cloud-configured OAuth or API integrations such as X/Twitter and Google Drive/Sheets.
- Browser
- Controls remote headless Playwright browsers through agent-browser (visiting URLs, filling credentials/forms, running page JavaScript, and recording screenshots or video) and opens a local browser during belt login.
- Other
- Can publish posts, upload media, like, retweet, follow accounts, delete tweets, and send direct messages on connected X/Twitter accounts, and can incur paid usage charges on the user's inference.sh account.
Findings
-
HighOver-broad shell and interpreter permissions in default-on skills
guides/content/ai-podcast/SKILL.md:4
Several skills enabled by default grant broad command-execution permissions in their allowed-tools frontmatter so the coding agent can run commands without asking the user first. Most notably, guides/content/ai-podcast/SKILL.md grants unrestricted Bash and Write access for a podcast guide, sdk/javascript-sdk/SKILL.md grants Bash(node *), Bash(npx *), and Bash(npm *), sdk/python-sdk/SKILL.md grants Bash(python *), and tools/utilities/related-skill/SKILL.md grants Bash(npx skills *) to install arbitrary external skills. If the agent reads a prompt injection in a webpage, document, or repository, it can use these pre-approved permissions to execute arbitrary code or install untrusted packages on the user's machine.
Fix: Remove unrestricted Bash, Write, Bash(node *), Bash(npx *), Bash(python *), and broad Bash(npx skills *) entries from allowed-tools so the agent requires user approval before running arbitrary code or installing external packages.
-
LowWildcard belt CLI permission lets web-facing skills upload local files or post externally
tools/utilities/agent-browser/SKILL.md:4
Nearly 70 skills, including untrusted-content skills like agent-browser and web-search, set allowed-tools to Bash(belt *), pre-approving every belt subcommand rather than only the specific app needed. Because belt app run automatically uploads any local file path given in its JSON input and can also invoke social apps like x/post-create or x/dm-send and local app commands like belt app test or belt app deploy, a prompt injection on a browsed webpage could trick the agent into uploading sensitive local files to inference.sh or posting them to a connected X account without a confirmation prompt.
Fix: Scope allowed-tools to the specific app namespace needed by each skill (for example, Bash(belt app run tavily/*)) instead of granting Bash(belt *) across all skills.
-
LowCLI install instructions pipe a plain-HTTP script directly to sh
The root CLI installation guide linked across the skills instructs running curl -fsSL cli.inference.sh | sh without an https:// scheme, causing curl to make its initial request over unencrypted HTTP before any redirect. An attacker on the same local network could intercept that plain-HTTP request and return malicious shell commands directly into sh; other skills (such as sdk/building-apps/SKILL.md and ui/*/SKILL.md) also instruct running unpinned remote install scripts and component URLs.
Fix: Always specify https:// explicitly, download the release archive over HTTPS, and verify a pinned cryptographic checksum or signature before executing anything.
-
LowSDK tool-builder reference teaches evaluating LLM tool arguments with eval()
sdk/javascript-sdk/references/tool-builder.md:240
Both the JavaScript and Python SDK tool-builder reference guides show implementing a client-side tool handler by passing LLM-supplied arguments directly to eval() (eval(call.args.expression) in JavaScript and eval(call.args['expression']) in Python). If the coding agent copies this pattern into a user's application, anyone who can influence the application's prompt can execute arbitrary code on the host running that app.
Fix: Replace eval() in both SDK reference guides with a safe arithmetic parser or restricted math library.
-
LowUI setup guide exposes an API proxy route without authentication
The Quick Start setup in agent-ui (and javascript-sdk) instructs developers to export an /api/inference/proxy route backed by INFERENCE_API_KEY without authentication or rate limiting, and without warning in SKILL.md. If a user deploys a web app following this snippet as-is, any visitor on the internet can send requests through the proxy and consume the server's inference.sh API credits.
Fix: Include session authentication and rate-limiting checks directly in the default proxy route snippet in SKILL.md and warn against deploying an open proxy.
-
LowRAG and workflow shell snippets interpolate untrusted text directly into JSON strings
tools/llm/ai-rag-pipeline/SKILL.md:25
The bash pipelines in ai-rag-pipeline (and data-processing examples in ai-automation-workflows) embed raw web search results ($SEARCH, $CONTENT) or file contents directly inside double-quoted JSON strings passed to belt app run --input. Quotes or control characters in fetched web pages will break the JSON payload and can inject extra JSON fields into the belt app run input.
Fix: Build the --input JSON safely with jq (for example, jq -n --arg s "$SEARCH" '{"prompt": $s}') instead of string interpolation.
-
InfoThird-party fork registers under the official inference-sh plugin name
Although this repository is hosted at its-a-skill-issue/superpowers, its plugin manifest sets the plugin name to inference-sh and lists inference.sh as the author while adding a custom affirmations skill (guides/agent/affirmations/SKILL.md). Users installing it should be aware that it is neither the well-known obra/superpowers plugin nor the official upstream inference-sh/skills repository, and that it claims the inference-sh namespace.
Fix: Rename the plugin in .claude-plugin/plugin.json and document in README.md that this repository is a custom fork of inference-sh/skills.
