ClawHub
Back to skill

Security audit

adversarial-debate

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed cross-vendor review workflow that sends a user-prepared brief to another model, with no hidden code or automatic repository scanning found.

Install only if you are comfortable sending the prepared debate brief, and possibly referenced source context depending on the configured CLI, to the selected external model provider. Redact secrets, credentials, customer data, regulated data, and proprietary material that should not leave your environment, and confirm the chosen backend before running a debate.

SkillSpector

Vulnerability Patterns
  • Rogue AgentSelf-Modification, Session Persistence
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (1)

Session Persistence

Medium
Category
Rogue Agent
Content
## What this skill sends where

Worth knowing before you run a debate: **the brief you write in step 1 is sent verbatim to a model hosted by another vendor.** That is the whole mechanism — there is no way to get cross-vendor adversarial value without it.

- The brief goes to whichever channel you selected (OpenAI through the Codex CLI, xAI through the Cursor CLI, or whichever backend you configured for handoff). Write it accordingly: it should carry the argument, not your secrets.
- Nothing else leaves your machine. This skill reads the brief you wrote and the attacker's replies. It does not scan your repository, collect telemetry, or send anything to a destination you did not configure.
Confidence
96% confidence
Finding
This skill explicitly instructs the operator to send a user-prepared brief verbatim to a third-party model hosted by another vendor. Even though the document warns not to include secrets, the mechanism creates a real data-exfiltration boundary: sensitive source content, internal plans, or regulated data could be disclosed externally if a user includes them in the brief or attached file paths/material. The skill context increases risk because cross-vendor transfer is the core behavior, not an incidental side effect.

Static analysis

No suspicious patterns detected.