The static analyzer found many command, environment, filesystem, and network patterns. Most high-volume detections are false positives from Markdown code fences, CLI flag tables, and official documentation links, but the skill does teach headless command execution with broad tools and CI secrets. No malicious intent or prompt injection was found, so publication is acceptable with security warnings.
Headless Claude Code Commands With Broad Tool Access
TRUE POSITIVE. The skill documents non-interactive Claude Code commands and examples that allow Bash, Read, Write, WebSearch, and auto-accepted edits. This is legitimate for automation, but unsafe prompts or untrusted repositories could cause file changes, command execution, or data exposure.
The examples clearly enable headless execution with powerful tools. The risk depends on operator configuration and prompt trust, so this is a confirmed contextual risk rather than malicious code.
Web Service Examples Invoke Claude Code From User Requests
TRUE POSITIVE. The Express and FastAPI examples accept request prompts and optional tool lists before spawning Claude Code. The examples use argument arrays, which limits shell injection, but they omit authentication, rate limits, prompt validation, and tool allowlists.
The spawn and subprocess calls are real and receive request-derived values. The code avoids shell interpolation, so the main concern is exposed automation control rather than direct command injection.
CI Examples Handle Repository Data And Secret Variables
TRUE POSITIVE. The CI examples combine pull request diffs, Claude Code execution, Git platform tokens, and ANTHROPIC_API_KEY. This is expected for CI automation, but users must avoid sending secrets or untrusted instructions to broad tool configurations.
The secret references are standard CI patterns and are not exfiltrated in the examples. Risk remains because untrusted PR content is fed into an automated agent.
Permission Bypass Mode Is Documented Without Guardrails
TRUE POSITIVE. The skill describes bypassPermissions as a mode that skips all prompts. In headless automation this can remove the last approval layer before tool use, especially when prompts include repository or CI data.
The cited lines directly describe bypassing prompts. No example uses bypassPermissions, which lowers severity, but publishing should warn users about this mode.
Markdown Backticks Misclassified As Ruby Shell Execution
FALSE POSITIVE. Many external command detections come from Markdown inline code, fenced shell examples, and CLI flag names. They are documentation examples, not executable skill code loaded by the marketplace.
The cited locations are Markdown code blocks and tables. There is no script file or runtime hook that executes these commands automatically.
Weak Crypto Detections Are Table Header False Positives
FALSE POSITIVE. The reported weak cryptography lines match words such as description or metadata, not cryptographic API usage. No evidence found of MD5, DES, SHA1 hashing, or encryption logic in the skill files.
Manual review found documentation text at the flagged lines, not cryptographic code. The scanner likely matched substrings inside ordinary words.
Official Documentation Links Are Benign Network References
FALSE POSITIVE. The hardcoded URLs point to official Claude Code documentation. They do not collect data, execute code, or direct traffic to an unknown endpoint.
The network indicators are plain reference links. No fetch, curl execution, or telemetry code is present at those lines.
FALSE POSITIVE. The filesystem detections at stderr redirection are normal shell examples using /dev/null. They do not read sensitive files or traverse user directories.
The file path is used only to suppress command errors. It is a standard shell redirection pattern with no exfiltration behavior.