The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface identified within the multi-agent coordination and reporting workflow. This vulnerability allows content from one participating model to potentially influence the host agent's instructions.
Ingestion points: Model performance comments in score-manager.js, relayed inter-model messages in herald.js, and task descriptions in team-coordinator.js.
Boundary markers: The system lacks explicit delimiters or "ignore embedded instructions" warnings when processing or relaying messages between models.
Capability inventory: The skill can create directories and write files (e.g., reports, configurations) to the local workspace via scripts/init.js and utils/template-renderer.js.
Sanitization: No prompt-specific escaping or validation is performed on model-generated content before it is interpolated into templates or relayed to the host model.
[DATA_EXFILTRATION]: The skill manages AI provider credentials by requesting API keys from the user and storing them in local configuration files (.trae/config/providers.json). While no network exfiltration or unauthorized access to external sensitive files (such as SSH keys) was detected, the storage of credentials in plain-text JSON files is a notable behavior.

teamwork