Today everyone's talking about the availability problems, but GitHub has a much bigger problem which is that its way of modeling work is not aligned with how people need to work with agents. I think the social value of GitHub has always been really overblown, and now it's actually an Achilles heel.

Now that we know prompt injection is such a massive risk, it seems insane to allow members of the public to leave comments on issues and pull requests that your agents may see. Which means the way that open source interacts with the public needs to fundamentally change away from the open GitHub model and towards a model of trusted contribution only. Feedback from the public has to come through carefully monitored channels that we can be sure our agents are not gonna pick up and run with. Public comment has to be kept totally segregated from our trusted work queue, which is how many people treat GitHub issues. We just simply can't have people, including malicious actors, able to drop a comment any and everywhere they want. I'm honestly astonished that GitHub hasn't even taken the most basic step of allowing you to create a read-only public repository with no commenting functionality.

This is only a problem if you have agents working from public comments and issues.

Why would you do such a thing?