The Agent Skills Directory

Indirect Prompt Injection (SAFE): The SDK documentation describes methods for processing external data via AI models, which is an inherent surface for indirect prompt injection. \n
Ingestion points: Untrusted data enters the context through the input object in client.run and message content in agent.sendMessage as shown in references/files.md and references/streaming.md. \n
Boundary markers: The examples do not demonstrate the use of delimiters or specific instructions to ignore embedded commands. \n
Capability inventory: The skill allows for network communication with api.inference.sh and local file reading via readFileSync or automatic path-based uploads. \n
Sanitization: No specific input sanitization or validation logic is presented in the reference material. \n- Data Exposure & Exfiltration (SAFE): The reference material describes intended file-handling capabilities, such as uploading local files by path. No sensitive system paths are used in examples, and the behavior is consistent with the primary purpose of the SDK. \n- Credentials Security (SAFE): The documentation provides proactive security advice in references/server-proxy.md, instructing developers on how to use backend proxies to avoid exposing API keys in browser environments.

javascript-sdk