blunt ships Markdown instructions, JSON manifests, and three hook/plugin scripts that read a local SKILL.md file and print it. The scripts run only when you install the skill and opt into always-on with a flag file. Nothing executes on install; nothing makes network calls; nothing collects data.

Use GitHub's private security advisory ("Report a vulnerability" on the Security tab) rather than a public issue. Include the runtime (Claude Code, Codex, Pi, OMP, OpenCode, or another harness), the install route, and the file involved.

• A hook or plugin script reading or executing files outside the checked-out repository.
• Injection through the flag file path or the plugin root environment.
• A manifest that makes an agent execute anything at install time.

• Prompt-injection behavior of the ruleset itself; the skill instructs models, and models remain responsible for what they do with instructions.
• Vulnerabilities in the agent runtimes (report those upstream).