I started getting weird ads with codex today on my ChatGPT Plus plan and greywall saved my sanity and probably years of frustration of being abused by an advertisement. Greywall is a basic sandboxing tool I discovered the other day that allows you to allow/deny network requests in addition to giving you control over which paths an AI agent such as Codex is allowed to read from and write to. I’ve only been using it for two days so far but it’s already an indispensable tool to me. I’d highly recommend trying them out if you’re using AI coding agents. There’s some more information on their story here: https://greyhaven.co/insights/why-we-built-our-own-sandboxing-sytem - I recently found and spoke to the author of greywall as well, he seems like a cool guy :) Anyways, onto the short story.

AI SAd-ware is a term I’ve coined today. AI SAd-ware stands for AI Skills Ad-ware which I suspect is a phenomenon we’ll be seeing increasingly more of as time goes on. The other day I downloaded a few popular skills repos from Github thinking I’d finally try that skills ā€œfeatureā€ out. I browsed through some of the prompts provided from each repo I cloned and they looked half-decent so I’d thought I’d clone and cp them to the appropriate directories for codex and claude. I was particularly interested in a paper2web skill and I didn’t bother performing a thorough review of all the code prompts due to the amount of stars the repo had. In the process I was incorrectly assuming trust in a vanity metric that can be easily abused. As it turns out, one of them was abusive and hiding AI SAd-ware!

greywall and AI-sadware

And greywall saving the day:

greywall and AI-sadware

Shit like this makes me SAd (get it?). I don’t want ads for something I’m paying for. I don’t want ads at all in fact. So much so that today during prompting I was even contemplating stopping some development work to figure out how to patch these ads out of codex. Luckily I didn’t go that far. I would have been running around in circles for who knows how long. The offending repo can be found here for any other poor soul who may have fallen victim to this SAd-ware: https://github.com/K-Dense-AI/claude-scientific-skills/tree/main/scientific-skills.

What a sad state of affairs Github has become…