The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains no instructions designed to bypass agent safety filters or override system constraints. It is focused entirely on task decomposition and multi-agent coordination logic.
[DATA_EXFILTRATION]: No patterns of data exfiltration or access to sensitive local files (like SSH keys or AWS credentials) were found. The 'Proactive Mode' performs legitimate local project scans (e.g., git status, git log, npm audit) to recommend development tasks, which is standard for an orchestrator in this domain.
[REMOTE_CODE_EXECUTION]: No unauthorized remote code execution patterns or downloads of unverifiable scripts were detected. The skill spawns specialist agents using platform-standard tools such as Claude Code's Agent and Codex CLI's spawn_agent.
[COMMAND_EXECUTION]: Shell command usage is limited to version control (git), project health checks (npm audit), and environment management within the context of developer tooling. No destructive or hidden commands were found.
[INDIRECT_PROMPT_INJECTION]: As a meta-orchestrator, the skill has a large attack surface by ingesting data from multiple agents. However, it implements robust mitigations, including mandatory schema validation for all inter-agent messages, structured handoff protocols, and an 'Evaluator Loop' (Pattern H) where independent agents verify deliverables against a contract before completion.
[PRIVILEGE_ESCALATION]: The skill utilizes platform-specific features like mode: bypassPermissions when spawning sub-agents. This is a standard method on the Claude Code platform for enabling autonomous multi-agent execution and does not represent a security bypass outside the intended platform functionality.

nexus