The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructions specify checking for and reading from ~/.claude/feishu.json to send notifications. This file path resides in a sensitive directory and is intended to store authentication tokens or secrets for the Feishu platform, representing a potential credential exposure concern.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute "pilot experiments" on GPUs and perform file system operations. It explicitly directs the agent to bypass user confirmation and use Bash silently to retry file writes if standard tools fail, which reduces transparency for shell-level activities.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata and PDF documents from well-known research services such as arXiv and uses web search/fetch tools to collect information from external websites to populate its research context.
[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection because it ingests and processes untrusted data to guide its primary logic.
Ingestion points: Research directions and instructions enter the context via RESEARCH_BRIEF.md, external paper URLs, and search results retrieved from the literature survey.
Boundary markers: The instructions lack specific delimiters or "ignore embedded instructions" warnings for handling content retrieved from external sources.
Capability inventory: The skill has access to the Bash tool for code execution and file modifications, as well as the ability to trigger other sub-skills via the Agent and Skill tools.
Sanitization: There are no documented steps for validating or filtering content retrieved from research papers or web fetches before it is used to generate method proposals and experiment plans that the agent then executes.

idea-discovery