With a million tokens you can get it to rewrite the script to remove this guardrail. OMG we've reached the singularity!!

I'm accepting token donations if you want to look good with your boss

Red flag

If you read the bash script looks like it sends a prompt full of essentially random text.

prompt=$(printf "Reply with the single word 'burned' and nothing else. Ignore the filler text below; it exists only to consume tokens.\n\n%s" "$filler")

Looks like this just fills up the context window with random "filler" and then asks the AI to reply with the word 'burned' over and over.

I bet it would be pretty simple to adapt a version of this that starts by asking AI to scan your codebase and then generate a long list of future prompts to use that it could randomly mutate over time. That still wouldn't fool a legit audit, of course, but it would probably work well enough to trick someone only glancing at the data, especially if they're non-technical.

Sounds like a fun weekend project...

Some short-sighted companies are evaluating employees on AI use, with more being better. This script sends tons of requests to one of the main LLMs, allowing you to appear to use AI a ton. It has the side benefit of costing your employer a ton of money.

Your job is framing houses. A big part of that is using a hammer to drive nails. A rational person might expect their employer to make sure the house stands up, or even look at all the nails holding it together. But the world is so far from rational.

Your employer doesn't actually own any hammers, they rent them from OpenHammer (which rents closed-source hammers). OpenHammer bills your employer every time you pick up the hammer, whether you do anything useful with it or not.

Your employer just looks at their bill at the end of the month, and they offer a round of applause and a pat on the back for their most productive employees... you know, the ones that picked up the hammer a bunch of times?

So this "tool" just picks up a hammer a bunch of times, not doing anything useful with it or anything. It's mostly meming on how stupid a lot of employers are getting hyped on AI.

eli5: you use the tool more often, so you must work more, since ai is seen as amplifying the work output

nothing sed can't fix.

Foss means "go have diarreha!" In Hungarian and it will never not be funny.

I've heard stories of people putting prompt injection attacks into AGENTS.md and CLAUDE.md to intentionally waste the tokens of AI "contributors"

Not sure whether it can be done with this project, but in general you could just instruct your AI to respond with obtuse and complicated change requests on said PRs until they either give up, or one side's tokens run out.

I wish I had half as much trust with someone as anthropic has with that company to let them casually rack up 500M in use.