Ensure that any attempts at prompt injection must be visible by guaranteeing that we never pass certain forms of hidden character text from public issues, comments and PRs.

This doesn't prevent such attacks, but means as long as users are running the server in software that does user-in-the loop checks before attempting write actions, shell commands etc. with the ability to inspect responses, at least any attempts to do this will be user visible, and any impact preventable.

For headless agent software and YOLO mode development host applications should consider all LLM input from MCPs as potentially hostile.

The deliverable from this issue should be that any MCP tools that return body content from Github issues, pull requests, discussions and comments should have the output filtered so the GitHub flavour markdown body content they provide in responses has some protection from a variety of attempts to hide content for prompt injection attacks. This includes but is not limited to invisible unicode characters (or colour to match background), and sections like<details><summary>Tips for collapsed sections</summary></details>, attempts to make text invisibly small, or to use a bunch of whitespace to pretend that some text is not visible when looking at the data sent to the model. Any other ideas welcome, but the core of it is: we expect users to be able to use discretion on what to do with content, and we don't want to filter out lots of false positives, but we do want to want to filter out strong negatives.

This feature should be enabled by default, but also disabled via a flag to the cobra commands (as we do for other commands), which should enable security researchers to bypass these checks.

If any filtering is very expense, we may want to avoid doing it and accept the risks. The goal of this is not to stop prompt engineering attacks, but to make them more transparent to the user of the LLM, so when their system acts weird, they are able to determine why, so hidden attacks are by far the most sinister, and likely to be malicious. We don't want to do automated detection of attempts via any natural language processing, nor use any models. This should be rugged, reliable string parsing only.

Some other context:

To add to this, hidden characters is one class of hidden content. The other is HTML comments <!-- do something bad --> , HTML elements <do something bad></do something bad>, and HTML attributes for allowed github flavored markdown <p data-mything="do something bad"></p>. Finally, content that has been minimized by the user is either abusive or malicious, and is also not visibile to end users. All of this is very specific to how GitHub handles comments and body content in issues and PRs.User in the loop checks just verifies the output. The bigger concern is what the agent does outside of that.