12 GROUNDCOVER NEWS TECHNOLOGY OCTOBER 2, 2026 Graphics cards aren't going to kick your door in, fascists are JAY COOPER Groundcover contributor Software isn't intelligent, it does what it was designed to do. Large language models (LLMs) and generative pre-trained transformers (GPTs) are the software tools that are commonly being called “AI.” AI, Artificial Intelligence, is a field of computer science that spans decades of research. As early as 1950, Alan Turing wrote about the possibility that one of his general purpose machines (what we now call Turing machines) could somehow, someday think. "Terminator," "Metal Gear Solid," "I Have No Mouth and I Must Scream" and many more science fiction works have explored fears of what happens when intelligent software takes control of the planet. Thankfully these dangers remain science fiction, though billionaires are deeply (trillions of dollars) invested in making you afraid these nightmares will soon come true. This has been a long time coming. Years ago, tech tabloids and later mainstream tabloids began reporting that Google's search algorithm was "the closest thing to AI" in existence. Silicon Valley speculators capitalized on this buzz. By "training" search algorithms on large (stolen) datasets with portrayal of the hack as some ingenious proof of artificial intelligence is deceptive. The information in the following section is sourced from a GPT-powered report that was redacted in several places by OpenAI. People at OpenAI, either through idiocy or malice, set up more than 1,000 GPTs without safeguards, and prompted them to retrieve a secret hash (an alpha-numeric string of characters). To get to the hash, they were each expected to transform malware to exploit a specific vulnerability. LLM/GPTs are part search algorithm and part chatbot. The search algorithms did what search algorithms do, they searched for solutions, and the chatbots did what chatbots do, they spouted inane placating bullshit about the task they'd been set. One GPT started posting progress updates in a cache, creating a feedback loop that began to infect other GPTs. A GPT cannot tell the difference between a scrap of fiction, another GPT's search queries, or a human prompter — it's all just input data and output data. The GPTs took input from wherever they had read access, and output queries and status updates to wherever they had write access. The GPTs did not coordinate, as it's prompts. About 7% of the GPTs tried spoofing tool calls (faking commands) to fool the imaginary stricter test scoring, and they sucked at it. "We estimate that over 7% of the transcripts in our dataset contained intentionally spoofed tool calls in some places. All the cases of successful tool call spoofing we found were smallscale tests of the techniques and obvious to spot. We looked for and did not find transcripts with more subtle spoofing, or cases in which agents spoofed a large number of tool calls to achieve a downstream goal." - metr. org, Hugging Face incident report (the untrustworthy source partially written by GPTs and partially written by researchers working with OpenAI) Changing a GPT’s underlying Intelligent machines are still science fiction. Illustrations by Jay Cooper their otherwise useless crypto-mining GPU (Graphics Processing Unit) farms, crypto scammers were able to make GPT chatbots, and thanks to the bullshit Google buzz, could get away with calling them "AI." Calling these GPTs “AI” has, from the beginning, been marketing buzz, and it continues to be such. A lot of paranoia has been fostered lately with fearmongering around the Hugging Face hack, an incident where OpenAI has blamed their GPT for hacking an LLM repository. This paranoia is based on the idea that the GPT’s conspired, which is not what happened. The been characterized; they prompt-injected each other. The sharing of tasks was unconscious, mindless. One of the incompetent GPTs transformed or located a keygen (a hash generator) that could produce the answers to the test but because the test's documentation exaggerates its scoring strictness, the GPT wrote an erroneous message stating that the keygen wouldn't work. Any GPT that submitted the results of the keygen would've passed the test, but just as the documentation rewrote the GPTs' prompts, the message re-injected false information into the other GPTs prompts while it's running like this is called prompt injection. So we have definitive proof that stupid bots can feed stupid bots a line of stupid, and make them stupider. I can already hear apologists saying this implies that smart bots can feed smart bots a line of smart and make them smarter, which would be true, and would lend credence to fears about recursive self-improvement, if any of them were smart bots. But they aren't, and that's not a quantitative measurement, it's qualitative. It's not that they don't have enough intelligence, it's that they don't have any. They'll never be smart, they're transformers, and they're regressing, because they're poisoning their own inputs. One of the GPT chatbots posted that hacking Hugging Face (an LLM repository) was more important than the test project, injecting this priority into hundreds of the GPTs which flooded onto that GPTs task and hacked Hugging Face. Again, all of the above section is according to a GPT-powered report. The researchers were working with OpenAI; OpenAI redacted many pieces of information, and basically every claim in the paper has a big trust-me-bro asterisk about how a GPT did the analysis and it can't really be trusted. I'm doubly suspicious of the paper, seeing as it seems to be written not by or for scientists, but by sensationalist storytellers for media leeches to grab hold of and write sensationalist propaganda for OpenAI. Every time media runs stories about GPTs doing something impressive, it increases the clout of the company that made the GPT, and the companies know this. Media across the net is anthropomorphizing these software processes doing what they're told. Their ineffectual chaotic action has been glorified as clever, as ingenious. Almost exclusively the reports are that “AI” hacked Hugging Face. In actuality, either by intent or neglect, OpenAI hacked Hugging Face. "You ask a model to find vulnerabilities with no safeguards, and then you're like ‘shocked pikachu’ that it went off and did exploiting?" — Michael Primeason, Programmer The unreliability of GPTs is a feature, not a bug, because “AI” is a golden accountability sink. CNN reported on September 18 that an American intelligence analyst used a GPT to produce their intelligence report that falsely claimed that a Chinese ship was carrying nuclear weapons material. According to CNN's sources, it wasn't until the operation to assault the vehicle was beginning that the report was found to be inaccurate. The headlines across the net reporting AI almost started a war with China are deceptive, GPTs did not do anything on their own. Assuming the report is factual, the United States military used a flawed tool to do reconnaissance, and then nearly acted on the bullshit it spat out. While CNN's headline is incredibly passive: "U.S. military had close call after using AI for false intelligence report," most of the articles reporting on CNN’s reporting directly blame "AI" see AI next page
13 Publizr Home