The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill prominently features instructions to download and execute shell scripts from an external GitHub repository belonging to an individual user, which is a significant security risk for arbitrary code execution.
Evidence: curl -fsSL https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.sh | sh found in SKILL.md and references/installation.md.
Evidence: irm https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.ps1 | iex found in references/installation.md for Windows users.
[COMMAND_EXECUTION]: The skill requires the execution of several CLI commands that interact with the local file system and network to install and manage the underlying engine.
[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and indexing local source code, which is then provided to the agent. This creates a vector where malicious instructions embedded in project files could influence the agent's behavior.
Ingestion points: Tools such as codegraph_explore, codegraph_search, and codegraph_node read contents of files in the current workspace.
Boundary markers: There are no explicit instructions for the agent to sanitize or ignore instructions embedded within the source code data it retrieves; conversely, the skill instructs the agent to "trust the returned results."
Capability inventory: The agent uses the retrieved code to explain architecture, analyze impacts, and guide coding tasks, allowing potentially malicious code comments or strings to hijack the conversation.
Sanitization: No sanitization of the source code content is performed before presenting it to the agent.

codegraph