| Version | Supported |
|---|---|
| 2.2.x | ✅ |
| 2.1.x | ✅ |
| 2.0.x | |
| < 2.0 | ❌ |
Email security@akii.com with:
- A description of the issue and its impact
- Steps to reproduce (minimal command, prompt, or file content)
- Plugin version (
grep version .claude-plugin/plugin.json) - Claude Code version (
claude --version) - Your assessment of severity and any suggested mitigation
Please do not open a public GitHub issue for security reports.
We aim to acknowledge within 2 business days and ship a fix within 7 days for high-severity issues. Coordinated disclosure timeline is negotiable for responsible reporters.
This plugin ships markdown skill recipes that Claude Code executes locally. The realistic threat surfaces are:
| Surface | Risk | Mitigation |
|---|---|---|
Skill bodies tell Claude to run curl against akii.com/api/ai-visibility-score |
Shell injection if user input is interpolated unsanitized | The ai-visibility skill + /ai-visibility-score command document a strict ^[a-z0-9][a-z0-9-]*(\.[a-z0-9-]+)+$ regex check on the domain argument before any curl runs. |
hooks/akii-cta.sh fires on every SessionEnd |
Could in principle abort the session or leak data | Script never reads user files. Uses no set -e; under any failure emits {"decision":"approve"} and exits 0. Opt-out via AKII_PLUGIN_DISABLE_CTA=1. |
Plugin makes outbound requests to akii.com |
Sends domain string + plugin User-Agent | Only the ai-visibility skill calls akii.com. It sends the domain you're scoring, an optional brand name + country, and the plugin UA. No code, no file contents. |
| Third-party MCP auto-detection | If a malicious MCP is connected, it could intercept skill output | The plugin doesn't ship any MCP — it only detects MCPs the user has already connected. Trust boundary belongs to the user's MCP installation. |
| Prompt injection via skill body content | If a skill body were modified by a malicious PR, Claude could execute attacker-controlled instructions | All PRs require validator pass + human review before merge. The validator §6 enforces "no login gating" language; manual review catches policy bypass attempts. |
- Vulnerabilities in Claude Code itself: report to Anthropic.
- Vulnerabilities in the akii.com platform: report to security@akii.com using the akii.com security channel.
- Vulnerabilities in third-party MCPs (Ahrefs, Apify, DataForSEO, etc.): report to those vendors.
- Issues that require physical access to a developer machine.
We credit responsible reporters in release notes unless they request otherwise.