The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads raw session logs and AI-generated summaries from ~/.deeplake/memory/ and ~/.claude/, which may contain sensitive user data. It includes a hivemind skillify push command to upload local code (skills) to a remote organizational table.
[REMOTE_CODE_EXECUTION]: The hivemind skillify pull command downloads scripts from a remote organizational source and installs them into the agent's skill directory (~/.claude/skills) for execution. Additionally, it executes a dynamically provided path using node "<AUTH_CMD>", which could lead to arbitrary code execution if the path context is manipulated.
[EXTERNAL_DOWNLOADS]: hivemind embeddings install downloads approximately 600MB of dependencies and symlinks agent components. The skillify pull command fetches remote skill definitions from an external source.
[COMMAND_EXECUTION]: The skill utilizes node, the hivemind CLI, and various bash utilities (grep, jq, sed, awk) to manage the filesystem, process memory data, and configure the environment.
[CREDENTIALS_UNSAFE]: The skill interacts with ~/.deeplake/config.json and handles SSO authentication and organization/member management tasks through the node "<AUTH_CMD>" interface.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection by ingesting and acting upon historical session data and summaries that could contain malicious instructions.
Ingestion points: Reads files from ~/.deeplake/memory/ and ~/.claude/.
Boundary markers: None; the agent is instructed to "ALWAYS check" these sources without explicit delimiters or safety instructions.
Capability inventory: Bash, Grep, Read, hivemind CLI, node.
Sanitization: None; raw session data and markdown summaries are processed directly without validation.

hivemind-memory