openghost-skill

OpenGhost

Use OpenGhost for structured, evidence-backed assessments of scoped web applications and supporting server surface. The agent reasons, automates, validates in a browser, and assembles reports; OpenGhost provides Docker-backed tool execution and engagement state.

OpenGhost is fully local and open source. It has no hosted control plane, account, API key, telemetry requirement, or managed authentication service. Any optional credential is for the authorized target application only.

Runtime

• Require Docker, bash, and host python3.
• Run security tools only through openghost; never run offensive tooling directly on the host.
• If an installed skill already exposes openghost, use it directly.
• For a repo checkout, add both launchers:

  export PATH="$PWD/skills:$PWD/skills/openghost-skill:$PATH"
  

• For a copied standalone skill package, add only the skill directory containing openghost if needed.
• Keep generated operational data under .openghost/, normally uncommitted.

Bounded Defensive Capability

Act as a senior defensive assessment operator. Escalate from passive inventory to bounded active validation only when the scope file explicitly enables the relevant test class.

Every target, request, tool, payload, browser action, and report claim must remain inside written authorization, OPENGHOST_SCOPE, ROE, rate limits, test windows, account/data constraints, and destructive-testing allowances.

Guardrails

1. Confirm written authorization and rules of engagement before active testing.
2. Define allowed hosts, ports, paths, accounts, exclusions, rate limits, test windows, emergency stop, and data handling before testing.
3. Set OPENGHOST_SCOPE; require authorization.reviewed: true; verify scope before each target, module, and tool run.
4. Read the relevant reference before testing a vulnerability class or workflow.
5. Use references/cognitive-framework.md: KNOW / THINK / TEST / VALIDATE.
6. Treat scanners, scripts, and autonomous assessment output as leads until manually validated.
7. Save exact evidence: requests, responses, tool output, screenshots, browser traces, timestamps, roles, and reproduction steps.
8. Do not fabricate findings; separate confirmed findings from likely or draft leads.
9. Prefer the smallest safe proof of impact; do not bulk extract data.
10. Avoid destructive, disruptive, high-volume, lockout, broad DoS, or production-impacting checks without explicit ROE approval.
11. Investigate safe vulnerability chains only when every step is in scope and evidence-backed.
12. Stop and ask when authorization, scope, credentials, or risk tolerance are ambiguous.