The Agent Skills Directory

[Prompt Injection] (SAFE): No malicious instructions, jailbreak attempts, or system prompt extraction patterns were found. The provided templates encourage structured task execution with explicit constraints (e.g., 'Do NOT modify any files. Research only.').
[Data Exposure & Exfiltration] (SAFE): No hardcoded secrets, API keys, or attempts to access sensitive system paths (like ~/.ssh or ~/.aws) were detected. The tools used are standard for development environments.
[Indirect Prompt Injection] (SAFE): The skill possesses a data ingestion surface through WebFetch and WebSearch tools, but it is not classified as a vulnerability here because the orchestration logic is focused on internal codebase management and includes instructions that limit the scope of subagent actions.
Ingestion points: WebFetch, WebSearch tools (reference/task-tool-guide.md).
Boundary markers: Templates use clear role-based instructions to delimit tasks.
Capability inventory: Edit, Write, and NotebookEdit tools are available to general-purpose agents for implementation tasks.
Sanitization: Not explicitly mentioned, but the 'Review Team' pattern is designed to provide human-in-the-loop style validation of changes.
[External Downloads] (SAFE): The skill depends on other logical agent skills ('extension-authoring-skill', 'agent-teams-skill') but does not perform any unauthorized remote script execution or third-party package installations.

subagent-teams