The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes ambient audio transcriptions, which are untrusted external data sources. The 'Deep Learning' workflow instructs the agent to extract facts, interests, and commitments from these transcripts to update a persistent user.md profile. An attacker could potentially inject malicious instructions via spoken conversation captured by the wearable device.\n
Ingestion points: Conversation data is retrieved via bee now, bee conversations get, and bee search commands.\n
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands within transcriptions are defined.\n
Capability inventory: The skill utilizes file-writing (to user.md and .bee-cursor), local file synchronization (bee sync), and command execution (bee CLI).\n
Sanitization: The instructions do not specify sanitization or validation of the transcribed text before it is used to update profiles or suggest todos.\n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of the @beeai/cli package and provides a link to download binaries from the bee-computer/bee-cli GitHub repository. These resources belong to the skill's authoring entity.\n- [COMMAND_EXECUTION]: The skill's primary functionality is implemented through the execution of the bee CLI tool to manage highly sensitive user data including transcripts, locations, and photos.\n- [DATA_EXFILTRATION]: The bee stream command includes a --webhook-endpoint parameter that allows streaming data to an external URL. If misconfigured or used maliciously, this could be used to exfiltrate private conversation data.

bee-cli