The Agent Skills Directory

[Data Exposure & Exfiltration] (SAFE): The skill follows security best practices by instructing users to store credentials in environment variables or .env.local files and utilizing lk app env -w for secure setup. No hardcoded API keys or secrets are present.
[Unverifiable Dependencies & Remote Code Execution] (SAFE): Installation steps use verified, official LiveKit packages and common utilities (e.g., python-dotenv). No unversioned dependencies or piped remote execution commands (e.g., curl | bash) were identified.
[Prompt Injection] (SAFE): Prompt examples provided in the Agent class instructions are benign and aligned with the functional purpose of a voice assistant. No bypass or safety filter override patterns were found.
[Indirect Prompt Injection] (SAFE): While the framework processes untrusted user audio via AgentSession, this is the primary intended purpose of the SDK. The examples do not demonstrate unsafe interpolation or handling. Application-level security is left to the developer.
[Obfuscation] (SAFE): No Base64, zero-width characters, or homoglyphs were used to hide instructions or code.

agents-py