The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install and execute an MCP server directly from a third-party GitHub repository (https://github.com/svnscha/mcp-windbg) using uvx. This source is not part of the trusted vendors list and is not explicitly linked to the skill author (novotnyllc), creating a risk of running unverified code.
[COMMAND_EXECUTION]: The live attach workflow (references/live-attach.md) involves launching cdb.exe with a TCP server listener (-server tcp:port=5005). This exposes the target process to the network, potentially allowing unauthorized access to the debug session if not properly restricted to localhost.
[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection through the data it processes.
Ingestion points: Crash dump files (.dmp) and live process memory are read into the agent's context via WinDbg commands (e.g., !pe, !clrstack).
Boundary markers: The instructions lack delimiters or warnings to ignore potentially malicious instructions embedded in the string data or exception messages found within the debugged process.
Capability inventory: The agent has the ability to execute arbitrary debugger commands (mcp_mcp-windbg_run_windbg_cmd) and generate diagnostic reports based on this untrusted data.
Sanitization: There is no documented process for sanitizing or escaping content extracted from memory before it is processed by the LLM, which could allow strings inside a crash dump to influence agent behavior.

dotnet-debugging