The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to send project context, including the "shared working tree" and "subject payload" (which may contain sensitive code or documents), to external AI providers such as OpenAI (Codex), Anthropic (Claude), X (Grok), and Google (Gemini). While the skill documentation claims this is an authorized read-only check, it constitutes a deliberate data egress path for repository content.
[REMOTE_CODE_EXECUTION]: The shell script scripts/cross-model-pov.sh uses npx -y to download and execute the acpx (Agent Client Protocol) package from the npm registry at runtime. This allows for the execution of unverified remote code during the skill's operation.
[COMMAND_EXECUTION]: The Python script scripts/peer-job-runner.py acts as a detached process manager that executes arbitrary worker commands. It uses double-forking on POSIX and detached process flags on Windows to run shell scripts and model CLIs outside the agent's immediate supervisor window.
[EXTERNAL_DOWNLOADS]: In addition to the acpx package download, the skill's external-evidence-researcher.md subagent is designed to fetch content from the public web using search and fetch tools, introducing external data into the agent's reasoning context.
[INDIRECT_PROMPT_INJECTION]: The skill utilizes "scouts" (subagents) that ingest untrusted data from project files and web search results. This creates a surface for indirect prompt injection where instructions embedded in project code or external documentation could influence the behavior of the agents or the final verdict.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes wrapper scripts (e.g., claude-safe-mode, pi-launch) to manage tool permissions and execution modes for various model CLIs at runtime.
[PRIVILEGE_ESCALATION]: The skill's documentation explicitly mentions the potential need for escalated permissions (require_escalated) in specific sandbox environments like Codex to allow the cross-model panel requests to reach external providers.

ce-pov