The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation encourages a 'Token-Efficient Pipeline' that uses save_response.py to write code generated by an external AI directly to the filesystem. This approach explicitly circumvents the standard human-in-the-loop or agent-based safety reviews (like the 'Write' tool), creating a direct path for executing potentially malicious code generated through manipulated AI responses.
[EXTERNAL_DOWNLOADS]: The tool described in the skill automatically fetches a binary WASM file (sha3_wasm_bg.wasm) from remote servers during the authentication process without explicit user verification.
[COMMAND_EXECUTION]: The skill provides extensive instructions for running local scripts and binaries (run.py, DeepSeekCode.exe) that have the capability to perform broad filesystem and shell operations.
[CREDENTIALS_UNSAFE]: The documentation specifies that session-sensitive data, including web bearer_tokens and cookies, are stored in a plain text config.json file, exposing them to any local process with read access.
[DATA_EXFILTRATION]: The skill's primary function involves sending local project data, including source code and configuration files, to external DeepSeek servers for processing.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing untrusted external inputs:
Ingestion points: Inputs via --delegate tasks, --agent goals, and --requirements documents (e.g., features.md).
Boundary markers: None identified in the prompt construction logic described.
Capability inventory: Full filesystem read/write and shell execution capabilities via integrated tools.
Sanitization: No input sanitization or validation of the requirements document is mentioned.

deepseek-code-mastery