If you find a security issue in ade-bench-plugin, please report it privately rather than opening a public issue.

• Email: security@typedef.ai
• Or use GitHub's private vulnerability reporting for this repo.

We'll acknowledge the report within a few business days and work with you on a fix and disclosure timeline.

This plugin is configuration for Claude Code — markdown skills, agent definitions, and small JSON/YAML manifests. It contains no Python or JavaScript runtime code. The interesting attack surface is:

1. The setup skill runs shell commands during install (git clone, uv sync, uv tool install, gh release download). Those commands are documented in skills/setup/SKILL.md and the user is the one who invokes the slash command. We don't ship binaries or execute remote code beyond those documented operations.
2. Generated tasks include unified-diff patches that are applied inside Docker containers by ade-bench at run time. The plugin itself does not apply patches outside the container.
3. The plugin reads files from your dbt project (model SQL, dbt_project.yml, profiles.yml) during task generation. It does not transmit them anywhere — Claude Code's context goes to whichever LLM provider you have configured, per your existing Claude Code privacy settings.

In particular, the plugin does not:

• Prompt for or store database credentials. The Snowflake setup step copies .env.example to .env for the user to fill in manually.
• Write to the user's source dbt project directory (only to the configured --output-dir).
• Modify shell config files, Docker state, or any system-level configuration.

• Vulnerabilities in ade-bench itself (the harness this plugin generates tasks for) — please report those at that repo.
• Vulnerabilities in Claude Code — please report those to Anthropic.
• Issues requiring an attacker to already have local code execution on the user's machine.

The plugin is at 0.1.0. While in this pre-1.0 phase, only the latest tagged version on main is supported. After 1.0, we'll publish a support window here.