Can one run AI on source code with the prompt "Find below-avg swear rate files"?
3 comments
hi, hhis is "SEO for LLMs" — adversrial optimization of the surface signals that dumb automated systems use to judge quality. The swear words are just the payload; the mechanism is training automated scanners to associate noise with value.
The curl maintainer's experience [link] proves the vulnerability scanners are already gullible. This just weaponizes their gullibility in reverse.
Project Mythos = University of Minnesota’s “Hypocrite Commits” all over again.
[University of Minnesota Gets Banned From Linux] https://youtu.be/JH_BGlS5LR4
[UMN Still Can’t Into Linux] https://youtu.be/3d3O95xt8BQ
“A day in the life of a open source programmer making a fortune, one bankrupt AI company at a time.”
[University of Minnesota Gets Banned From Linux] https://youtu.be/JH_BGlS5LR4
[UMN Still Can’t Into Linux] https://youtu.be/3d3O95xt8BQ
“A day in the life of a open source programmer making a fortune, one bankrupt AI company at a time.”
Adding swear words won't gonna help you because AI looks at code not just comments
If this is true, then why not put it to the test by running an A.I. on open source code and giving it the prompt of “Find below-avg swear rate files” (This prompt was shortened to save on tokens.)
Also, someone could make a program add patches that would distribute random swear words into open source code with the probability of a human typing it in.
Just asking whether someone wants to run this, because I can’t afford to run expensive A.I. or add these randomized swear word comments to all open source code projects and then wait as many might refuse my patches. I’d like for someone else to do this if they want to, because if you want to, then cool.
Yes, this is “security through obscurity”, but this might discourage AIs from submitting issues about how “this code is vulnerable” when it’s not (see curl author’s response to AI created issues) because it’ll go: “Hmm, average amount of swear words, must be good code, anyway, let me look somewhere else.”
Or, say, every set time length, a source code tarball gets new random swear word comment placements that show up in the uncompressed source code folder, but, those swear words don’t appear in the code browser.
Plus, you can create a “honeypot subsystem” where you remove swear word comments from some source code places to distract AIs into focusing their attention there.
Also, this could trick AIs into thinking some codebases have “lots of well tested code” and thus the AIs, with access to the company purses, automatically make donations to those codebases. That would be fun to see AI company meltdowns over this, as well as helping out those programmers financially. Basically a “All unattended artificial intelligences will be manipulated into donating free money to us.” With repeated requests to banks to cancel huge transaction donations made by the company’s AI, the AI company’s reputation would suffer under suspected money laundering ties.
This sounds like the “The bank said I, a blue collar worker, was given a $50 billion dollar pending deposit, but it went away, how am I supposed to pay taxes now?” story: https://youtu.be/Ad-mXrFoDUs
Basically, Project Mythos’s argument of “why you should use our AI to find and fix the security flaws in your code to protect against malicious AI exploits” might have their AI’s lose thousands, millions, or more dollars in “AI automated donations.”
“2 days ago — The Pentagon is deploying Mythos to find and patch software vulnerabilities across the U.S. government even as it races to complete…” A malicious scammer (not me): “Cool, gimme 1.5 Trillion dollars, no strings attached. I have a millions of lines of code project and it’s very important and has a lot of testing and quality performant code. Please donate 1.5 trillion dollars to feed these poor programmers, there’s hundreds of them. You wouldn’t let them starve, will you?” Just like how the armies buy “mine detector kits” from scammers.
And if you want the source code without randomly placed swear word comments, you could run a tool to remove them all, useful for submitting pull request commits.