Post B5PwegimEzDRmMUAHA by lucy@starlightnet.work
 (DIR) More posts by lucy@starlightnet.work
 (DIR) Post #B5PwegimEzDRmMUAHA by lucy@starlightnet.work
       1 likes, 2 repeats
       
       We gotta fight back. Contacting other open source projects and asking them to (hopefully) adopt a no-AI policy is a good first step, but we need to go further.We need to make it harder for users to use AI with our content, and we need to make it harder for these companies to steal our shit.I'm working on a tool, called cyanide, in order to poison LLMs at inference-level so they're absolutely useless when told to summarise content from a given website.I encourage you to do the same. figure out ways to poison and otherwise break LLMs when they deal with your content. At inference level, at training level, it doesn't matter.The more people do this, the more diverse tooling we build to stop these AI bro fucks from stealing our shit, the harder it is for the AI companies to clean up our shit, and the more obvious the failures and shortcomings of LLMs as search engines become for the average user.Here are a couple resources I found helpful when researching LLM security topics:- promptfoo.dev, a database of LLM vulnerabilities with links to research papers, which models are affected, etc.- [google scholar, since these things are an active field of research](scholar.google.com)- [openrouter.ai, cheap easy testing, especially when tested against multiple models since it's one simple API](openrouter.ai)feel free to ping me for resources to add, and other tooling to break LLMs with the data they scrape, parse or train on itself (so while iocaine is cool and helpful, it doesn't really poison inputdata itself, it feeds different data depending on who accesses the page)