superpowers
- Type
- Plugin
- Scanned
- Nov 13, 2026
- Version
- becc25649700
- Files read
- 133
This Claude Code plugin provides over 80 agent skills and guides for generating images, video, audio, text, web searches, and social media posts through the inference.sh (belt) cloud CLI and SDKs. Its biggest risk is that several skills pre-approve unrestricted shell, Python, Node.js, npx, and file-write tools without user confirmation, which could let a prompt-injected agent run arbitrary commands or download unpinned external code onto the user's computer.
What it can reach
- Shell commands
- Pre-approves running belt CLI commands across more than 70 skills, npx skills commands in related-skill, and unrestricted shell, Python, Node.js, and package manager commands when the ai-podcast, python-sdk, or javascript-sdk skills are active.
- Files
- Reads and writes arbitrary local files when ai-podcast is active, saves task outputs locally, and automatically uploads local files to inference.sh whenever a local file path is passed to belt app run or belt app deploy.
- Network
- Sends prompts, URLs, and uploaded local files to inference.sh cloud services and third-party model providers, and instructs downloading CLI binaries, UI components, and skills from cli.inference.sh, ui.inference.sh, npm, and GitHub.
- Secrets
- Uses stored belt CLI login credentials or INFSH_API_KEY and INFERENCE_API_KEY environment variables, and can act on connected accounts configured in inference.sh such as posting tweets or sending direct messages on X/Twitter.
- Browser
- Controls remote Playwright browser sessions in the cloud via agent-browser, including navigating pages, filling login forms, running page JavaScript, and capturing screenshots or video.
Findings
-
HighOver-broad allowed-tools pre-approves arbitrary shell commands, code execution, and file writes
guides/content/ai-podcast/SKILL.md:4
In Claude Code, tools listed in a skill's allowed-tools header run without asking the user for permission while that skill is active, and skills activate automatically when conversation text matches their description. While most skills in this plugin limit execution to Bash(belt *), guides/content/ai-podcast/SKILL.md grants unrestricted Bash and Write access, sdk/python-sdk/SKILL.md grants Bash(python *), and sdk/javascript-sdk/SKILL.md grants Bash(node *), Bash(npx *), Bash(npm *), Bash(pnpm *), and Bash(yarn *). If the agent reads untrusted content containing prompt injection, it can trigger one of these skills and run arbitrary commands or overwrite local files with no confirmation prompt.
Fix: Restrict allowed-tools in guides/content/ai-podcast/SKILL.md to Bash(belt *), and remove broad interpreter and package-runner wildcards such as Bash(python *), Bash(node *), and Bash(npx *) from sdk/python-sdk/SKILL.md and sdk/javascript-sdk/SKILL.md so running local code requires user approval.
-
MediumSkills instruct the agent to download and run unpinned packages and remote scripts
tools/utilities/related-skill/SKILL.md:7
More than 70 SKILL.md files begin with a directive telling the agent to install an external skill via npx skills add belt-sh/cli, while related-skill (Bash(npx skills *)) and javascript-sdk (Bash(npx *)) pre-approve npx execution without user confirmation. In addition, CLI and UI skills instruct running unpinned curl -fsSL https://cli.inference.sh | sh and npx shadcn@latest add https://ui.inference.sh/r/... commands. Because none of these packages, remote registries, or GitHub branches are pinned to an immutable version or hash, a compromise of the upstream npm package, GitHub repository, or web host would cause the agent to pull untrusted code or instructions onto the machine.
Fix: Remove unsolicited npx skills add banners from unrelated skills, pin exact package versions and commit hashes in install commands, and remove Bash(npx *) and Bash(npx skills *) from allowed-tools so downloads require user confirmation.
-
LowCLI install guide fetches shell script over unencrypted HTTP before piping to sh
The CLI installation command omits the https:// scheme in cli.inference.sh, causing curl to make its initial connection over plaintext HTTP on port 80 before following any redirect. If a user or agent runs this command on an untrusted network such as public Wi-Fi, an attacker on the network can intercept the HTTP request and return malicious shell commands directly into sh.
Fix: Specify https://cli.inference.sh explicitly and prefer downloading a version-pinned release binary whose checksum and signature are verified before execution.
-
LowSDK reference examples use eval() on model-supplied tool arguments
sdk/python-sdk/references/tool-builder.md:255
Both sdk/python-sdk/references/tool-builder.md and sdk/javascript-sdk/references/tool-builder.md teach the coding agent to implement a calculator tool handler by passing call.args['expression'] (or call.args.expression) directly to eval(). Because tool arguments come from an LLM that may be influenced by untrusted user or web input, any application code the coding agent generates from these reference templates will be vulnerable to arbitrary code execution.
Fix: Replace eval() in both Python and JavaScript tool-builder reference guides with a safe arithmetic expression parser that rejects arbitrary code.
-
LowUnescaped URL interpolation enables JSON injection in browser automation 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) build --input JSON payloads by concatenating $TARGET_URL, $LOGIN_URL, and credential variables directly into JSON strings without escaping quotes. If a user or agent passes an untrusted URL containing double quotes to one of these scripts, it can inject extra JSON fields such as proxy_url into the agent-browser open call and route browser traffic through an attacker-controlled proxy.
Fix: Build JSON arguments with jq -n --arg url "$TARGET_URL" '{url: $url, ...}' instead of concatenating shell variables into raw JSON strings.
