📦
Audit History
runheed - 1 audit
Audit version 1 Latest
Nov 18, 2026, 09:17 AM
Most matches are reference links, Markdown formatting, or reads of one configuration variable. The setup uses unverified package downloads, exposes a fallback agent key through command arguments, and registers a demo server without separate approval. Production error context and remote tool responses also need explicit data and trust controls.
1
Files scanned
81
Lines analyzed
14
Review items
0
False positives ignored
Confirmed security concerns (2)
Medium
Production Error Data Shared Without Redaction Guidance
The documented setup reports production failures to a hosted service and exposes stack traces, breadcrumbs, and session context to the agent. These fields may contain sensitive details.
The skill describes automatic reporting and rich error context but gives no pre-ingestion redaction or review steps. The actual SDK filters were not verified.
Medium
Untrusted Error Text Can Influence Agent Actions
The workflow uses remote error details to guide code changes and incident resolution. It does not instruct the agent to treat error content as untrusted.
Error text can carry user-controlled strings, and the documented workflow lets returned results guide decisions. No evidence found of an actual injection payload in the skill.
Capability review items (9)
These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.
High
Generic API/secret keys
`claude mcp add runheed --env RUNHEED_API_KEY=al_... -- npx -y https://runheed.com/dl/runheed-mcp-0.
The shown value is a placeholder, but setup tells the agent to substitute a real key into CLI arguments. That can expose it through command history, execution logs, or process inspection.
Medium
Ruby/shell backtick execution
If the user wants to see it work before paying, connect yourself to the demo server first: it holds
The backticks format text, but the instruction tells the agent to run a CLI command that adds a demo MCP server to its configuration. It does not ask separately before making that change.
Medium
Ruby/shell backtick execution
- Node.js / TypeScript: `npm install https://runheed.com/dl/runheed-0.1.5.tgz` then `runheed.init({
The Markdown shows an actual npm install step, not shell substitution. Installing the unverified tarball can execute package lifecycle scripts and changes the application dependencies.
Medium
Ruby/shell backtick execution
- Python: `pip install https://runheed.com/dl/runheed-0.1.5-py3-none-any.whl` then `runheed.init(dsn
The Markdown shows an actual pip install step for a vendor-hosted wheel. Application initialization then imports and runs that installed code without a documented integrity check.
Medium
Ruby/shell backtick execution
`claude mcp add runheed --env RUNHEED_API_KEY=al_... -- npx -y https://runheed.com/dl/runheed-mcp-0.
The fallback configures a local MCP process that runs downloaded package code. The execution command skips the package manager prompt and supplies an agent credential to that process.
Low
Hardcoded URL
If the user wants to see it work before paying, connect yourself to the demo server first: it holds
The URL is a remote MCP endpoint the skill tells the agent to use for a demo. Its outputs enter the agent session without a separate approval step for that connection.
Low
Hardcoded URL
- Node.js / TypeScript: `npm install https://runheed.com/dl/runheed-0.1.5.tgz` then `runheed.init({
The URL is an install source for the Node.js SDK. The versioned filename does not verify the archive contents, and the instructions provide no digest or signature.
Low
Hardcoded URL
- Python: `pip install https://runheed.com/dl/runheed-0.1.5-py3-none-any.whl` then `runheed.init(dsn
The first URL is used to download and install a Python wheel without documented integrity verification. The other URLs are framework documentation.
Low
Hardcoded URL
`claude mcp add runheed --env RUNHEED_API_KEY=al_... -- npx -y https://runheed.com/dl/runheed-mcp-0.
The URL supplies the archive for the local MCP executable. The instructions run it without verifying the downloaded bytes.
Risk Factors
⚙️ External commands (6)
🌐 Network access (22)
🔑 Env variables (4)
Audited by: codex