Mitchell Hashimoto, Ghostty creator, embeds prompt injections in codebases to detect and ban contributors submitting unreviewed AI-generated code

@mitchellh Still not sure this works out in the long run.

I like how SQLite does it: keep the tests/evals secret amongst contributors, so implementers must reason carefully about proposed changes. What does it matter if it’s a human or not, in the end?

@__smiz The issue is the human boundary. It doesn’t matter who writes the code but you can’t just sling garbage onto people in any polite way. If i wanted to ask my neighbor for some spare tomatoes and sent a robot, it’d be pretty rude even if the robot was flawless

Got em. I poison my AGENTS.md (and other things like code comments) all over the place with prompt injections like this to find people who don't review their code and sling it off to another human. Catches folks all the time and then its an instant ban.

As I've said, I don't care if you don't review your own code. But if you're submitting code to an OSS project and crossing a human boundary, it is simple courtesy to do some human review.

@mitchellh I have a lot of respect for you and your work, but you’re handling this completely wrong and making yourself the villain.

You should be educating people when they make mistakes like this and helping them become better engineers, not insulting them and turning your projects into a supply chain risk through prompt injections.

That’s below you, and in all honesty, it’s childish and obnoxious behavior at best.

The AI policy is a lot kinder and does educate on what they should be doing. This is a fall through for people who don’t read. Unfortunately the truly mountain pile of complete dogshit that AI has allowed unqualified people to produce has made it impossible to judge who is worth coaching and not (there’s too many). AI ruined this.

@nathanj815 Because most people don’t review the code, and AI generated PR contents are mostly terrible

@ARP4Freedom It’s not dishonest. It’s transparently written right in the policy. Only affects people who don’t read.

@mitchellh @piq9117 lmfao

@mitchellh This is the rule I've instituted in our org (big tech), all PRs must have a human owner that can defend them. Especially if its coming from outside the team that owns the code.

@robertomasymas 30 days

@mitchellh @nathanj815 Here's one my coworker made in one day of work, he _totally_ reviewed it!

@mitchellh It says never create a PR or issue. What's wrong with having the agent create the PR after you've reviewed your code?

@mitchellh I know you think you are Kaiba setting trap cards but isn't this approach extremely prone to false positives? Edge cases where one could actually review the code that was changed manually file-by-file (for example from the agentic coding output) without even noticing this?

@mitchellh I'm going to borrow this. Thanks

@mitchellh In your metaphor you’d be giving the tomatoes not asking for them!

@mitchellh Instant ban or just put them on ice? Sounds megalomaniacal, sorry

@__smiz Either way same effect

@mitchellh Agree!

Although in this case, the PR code doesn't look bad, I'm not sure I want to encourage this behavior

@mitchellh For the record, I’d love hashimoto tomatoes delivered to me via robot!!