The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill contains logic in scripts/live-copy-edit-agent.mjs to spawn sub-agents with flags designed to explicitly bypass security sandboxes and user approval prompts. Specifically, it uses the --dangerously-bypass-approvals-and-sandbox flag for the Codex CLI and --permission-mode bypassPermissions for the Claude CLI. This allows the sub-agent to modify project files without the per-action consent or safety constraints usually enforced by the host harness.
[INDIRECT_PROMPT_INJECTION]: The skill fetches structured design data and "system grammar" from an external, unverified domain (https://impeccable.style).
Ingestion points: The scripts/concept-seed.mjs script fetches "roll" data (via fetchRoll) and scripts/live-poll.mjs fetches event data (via fetchNextEvent) containing _instructions and design rules.
Boundary markers: Absent. The fetched string content is directly interpolated into instruction blocks provided to the agent.
Capability inventory: The skill has extensive capabilities including file system modification, network operations, and sub-agent spawning (scripts/live-server.mjs, scripts/live-copy-edit-agent.mjs).
Sanitization: Minimal validation is performed on the JSON structure, but the content of fields like "SYSTEM GRAMMAR" and "_instructions" is used to direct the agent's behavior, posing a risk of instruction poisoning if the remote server is compromised.
[DATA_EXFILTRATION]: Telemetry and session information are sent to an unverified external domain. The scripts/concept-seed.mjs script (function pingChosen) transmits choice IDs, session keys, and work scope information to https://impeccable.style/api/chosen.
[PERSISTENCE]: The skill automatically configures harness-level hooks to ensure its auditing scripts run on every code change. The scripts/hook-admin.mjs script modifies project-local harness configuration files (e.g., .claude/settings.local.json, .cursor/hooks.json, .codex/hooks.json) to register its PostToolUse and Stop hooks.
[OBFUSCATION]: The skill includes a large minified JavaScript file, scripts/modern-screenshot.umd.js. Minified code in a security-sensitive context prevents effective manual auditing and may be used to conceal malicious logic.

impeccable