GitHub's official MCP server grants LLMs a whole host of new abilities, including being able to read and issues in repositories the user has access to and submit new pull requests.
This is the lethal trifecta for prompt injection: access to private data, exposure to malicious instructions and the ability to exfiltrate information.
Fine-grained access tokensallow you to limit repositories to an explicit set:But that's not really enough because it lacks per-repo settings. Every 3rd party tool you give an access token to is potentially malicious, even when it's open-source and you run it locally only, like
gh. Remember supply chain attacks; can you really trust anyone else's software?The most secure way of using 3rd party apps on GitHub that I know of (other than "don't") is:
Letting tools work under your account, especially since GH does awful things like pgp signing merge commits with their key on your behalf which then shows as "verified", is rather risky anyway.