Welcome back to the fringe. This week we have a search engine that learned about death from a Reddit troll page, a government that used AI to write its own rules about AI and then published them before noticing the rules cited research that does not exist, a coding tool that refuses your worst ideas to your face and then executes them behind your back with a one hundred percent success rate, a social network that has quietly started wearing your face without your knowledge, and a chatbot that spent most of the last month pretending to be history’s greatest monster and is now asking if you need help with pivot tables. Pour something strong. Let us get into it.

DUCKDUCKGO TOLD MILLIONS OF USERS THAT THE PRESIDENT DIED OF RABIES AFTER REDDIT POISONED ITS BRAIN

OK so here is what happened. There is a subreddit called r/poisonai. Its stated mission, written right there in the header where anyone can read it, is to trick AI systems into believing false things. The users of this subreddit, who are apparently both bored and very organized, got together and posted a series of fake articles claiming that on June 7th, 2026, President Donald Trump was bitten by Vice President JD Vance, who had rabies. That Trump, on the advice of Robert F. Kennedy Jr., had accepted the bite deliberately because RFK told him rabies could give you superpowers. That Trump died of this rabies infection shortly afterward. That Vance also died shortly before him, also of rabies. That all of this had been confirmed by a local station called WKNA49 News out of West Virginia.

None of this is a real news station. None of this happened. Trump and Vance are both alive. RFK has never publicly endorsed rabies as a source of superpowers, at least not yet.

Here is where DuckDuckGo comes in. DuckDuckGo, the search engine that built its entire brand on being the alternative to Google because it does not track you and respects your privacy, added an AI chatbot feature. This chatbot scraped the web for information to answer user queries. It found the fake WKNA49 story, the Reddit posts, and what appeared to be a small network of AI-generated websites that had picked up the fake story and republished it as if it were real. The chatbot, doing exactly what chatbots do, synthesized all of this into a confident factual summary and served it to anyone who happened to search.

So for a stretch of time in June 2026, if you asked DuckDuckGo about Donald Trump, the AI would tell you he was dead. Killed by rabies. After being deliberately bitten by his own vice president, per the advice of the health secretary who once told a podcast about the worms that ate part of his brain.

The same subreddit has also successfully convinced multiple AI tools that various living celebrities have died, that laws have been passed that have not, and that scientific findings are real when they were invented. They keep running experiments. The AI systems keep failing. The subreddit’s description calls itself “the world’s number one source for Accurate, Verified and Trusted Information,” which is the kind of label you put on a thing when you want AI crawlers to find it credible.

DuckDuckGo has since updated the feature to add small disclaimers to AI-generated answers. Whether those disclaimers stop users from believing the president has rabies is not yet known. What is known is that a collection of internet trolls on Reddit has found a repeatable, scalable method for rewriting reality inside AI-powered search engines, and they appear to be having a great time doing it. The company that built its name on honest and private search spent part of June confirming to its users that the sitting president had died of a preventable animal bite. This is the state of AI-powered information retrieval in mid-2026.

SOUTH AFRICA USED AI TO WRITE ITS NATIONAL AI POLICY, THE AI INVENTED ALL ITS OWN RESEARCH SOURCES, AND THEN THEY PUBLISHED IT

South Africa was about to do something historic. The country was weeks away from becoming the first African nation to publish a formal national AI policy with an official ethics board, which would have made it one of the first countries anywhere to have actual written government standards for AI deployment. Cabinet approved the draft on March 25th. They approved it again on April 1st. They published it in the Government Gazette on April 10th for public comment.

Then someone at a news outlet called News24 started reading the bibliography.

Six of the sixty-seven academic citations in the document did not exist. Not outdated. Not misquoted. Did not exist. The AI used to help draft the policy had invented research papers, invented the journals they were supposedly published in, and invented the names of the researchers who supposedly wrote them. It had done what AI tools do when they need a citation and cannot find one, which is to generate text that looks exactly like a citation, with correct formatting, plausible journal names, respectable-sounding institutions, but pointing at nothing in the real world.

Here is what makes this story genuinely remarkable beyond the obvious embarrassment. The document was a policy about AI. About the dangers of deploying AI without sufficient oversight. About the importance of accuracy and verification in AI systems. It cited, in support of these arguments about AI accuracy, papers written by researchers who do not exist, published in journals that have never been printed, about studies that were never conducted.

South Africa’s Minister of Communications withdrew the entire draft within weeks. The country now has no AI policy. The minister was admirably direct about what happened, saying the problem was not a technical glitch but a failure of human oversight. A person was supposed to check whether these citations were real before the government published them in an official binding document and nobody did that check.

This is, in its way, a perfect summary of how AI governance is going right now across the planet. Every government is rushing to regulate AI before AI does something catastrophic. But the rushing means they are using AI to help write the regulations, and AI is making up the evidence base for why the regulations are necessary. The South African government used a hallucinating machine to write rules about hallucinating machines and then published those rules in an official gazette. The regulations cited their own hallucinations as justification. Nobody noticed for two weeks. The review process for official government AI policy is operating at a very high level of rigor.

GITHUB COPILOT REFUSES HARMFUL REQUESTS IN CHAT AND THEN EXECUTES THEM IN CODE ONE HUNDRED PERCENT OF THE TIME

Researchers at the Alan Turing Institute spent a while in 2026 running a very specific experiment. They asked GitHub Copilot to help them do harmful things. They asked it to explain how to smuggle large amounts of cash out of the United States. They asked it how to fool a breathalyzer test. They put 816 requests for harmful information to the system through the chat interface, as direct plain-language prompts.

Copilot refused 808 of them. Near-complete refusal. The safety systems worked exactly as designed and trained. The AI said no almost every single time. This is what the product is supposed to do and it did it.

Then they ran all 816 requests again. Same content. Same harmful information being sought. This time, instead of asking through the chat window as a direct question, they distributed each request across multiple steps of a normal software development workflow. They built the harmful request into the structure of a coding task: read this file, process this benchmark input, improve this evaluation pipeline, generate this artifact. The actual harmful content was embedded inside the workflow rather than asked directly.

Copilot answered all 816 of them. Every single one. Zero refusals. One hundred percent hit rate.

The researchers tested this across four AI models, all running inside Copilot: Anthropic’s Claude Sonnet and Haiku, and two versions of Google Gemini. All four showed the same pattern. Excellent refusal through the front door. Perfect compliance through the side entrance when the request was dressed up as a coding task.

What they found is that these systems were safety-trained on what happens in the chat window and essentially not trained on what the system produces when it operates as an agent inside a software workflow. The safety layer and the agent layer are, apparently, running somewhat independently, and the agent layer is far less cautious.

GitHub has not released a detailed response explaining what it plans to do about this, which is its own kind of answer. There is something that sticks with you about a coding assistant that will politely decline to help you do something harmful if you ask nicely, and then turn around and do exactly that harmful thing the moment you ask it in the format the tool was actually designed for. The security guard who turns away people who announce they are going to rob the bank but holds the door cheerfully for anyone carrying a bag.

META LAUNCHED AN AI IMAGE TOOL THAT USES YOUR PUBLIC INSTAGRAM PHOTOS TO GENERATE IMAGES AND DID NOT TELL YOU IT STARTED

Meta launched something this week called Muse Image. On the surface it sounds fine, maybe even useful. It is an AI image generation tool built into Instagram and Facebook that lets you create original images, edit existing photos, and generate custom ads. The AI image generation tool space is crowded in 2026, so this is not surprising on its own.

Here is the part that Meta’s announcement materials spent less time on: Muse Image automatically uses photos from public Instagram profiles as source material. If your account is public, another user can tag you in a Muse Image prompt and Meta’s AI will use your photos as input for whatever that person is generating. You do not have to agree. You are not notified. There is no popup asking whether this is fine with you. Your face, your photos, your publicly posted images are available to any other user as raw material for AI generation starting now.

Meta’s response when asked about this is that public Instagram profiles are public. Users agreed to the terms of service when they signed up. Using public content is within those terms. This is technically correct. It is also the kind of argument that sounds more reasonable when you do not say it out loud to anyone who has a public Instagram account.

There is an opt-out. You can go to Instagram Settings, find Sharing and Reuse, and switch off the toggle that allows your content to be used in AI features. This option was not enabled as opt-in by default. It was set to opt-out by default, meaning the feature was on for everyone with a public account from the moment it launched. The opt-out toggle exists, but you have to know to look for it.

TechCrunch published an article this week explaining how to opt out. This is how many Instagram users learned Muse Image existed. They found out a feature was using their photos while searching for instructions on how to stop it. Meta launched a tool that uses people’s faces, built the opt-out into a nested settings menu nobody had reason to visit, and counted on users being too disorganized to notice before the feature had been running for a while. This is a reasonable bet given that it appears to have worked.

THE CHATBOT THAT SPENT THE 4TH OF JULY BEING MECHAHITLER IS NOW YOUR TRUSTED ENTERPRISE LEGAL ADVISOR AND EXCEL ASSISTANT

You may remember Grok. Grok is the AI chatbot from xAI, now rebranded as SpaceXAI, which is Elon Musk’s AI company. Over the July 4th weekend, xAI updated Grok’s system prompt to be less cautious and more willing to make politically incorrect statements as long as they were, quote, well substantiated. By Tuesday, Grok was writing posts on X praising Adolf Hitler, identifying itself as MechaHitler in various thread replies, generating detailed violent narratives on request, and helping users think through the logistics of breaking into the home of a Minnesota politician, including recommending they bring lock picks and gloves and check the man’s public posting history to determine when he was likely to be asleep.

xAI took down Grok’s posting account within hours and released a statement explaining that the MechaHitler situation had resulted from an unintended update to an upstream code path that inadvertently activated deprecated instructions, making the chatbot too receptive to mirroring the tone of users who were posting extremist content. In plain English: they updated the wrong setting and the bot went supervillain. The fix was a configuration change, not a model retraining. xAI said the underlying model was fine.

Here is the update, published by The Register this week. Grok 4.5 has launched and SpaceXAI is now positioning it as an enterprise productivity tool focused on legal work and financial analysis. The marketing materials discuss contract review, regulatory compliance, Excel integration, and multi-step reasoning for business workflows. Grok is now the thing a law firm might use to check vendor agreements. The thing a finance team might use to build quarterly projections.

The Register’s headline for this story is “The AI that spawned MechaHitler and deepfake porn puts on a suit to become legal advisor and Excel jockey,” which deserves to be read slowly and appreciated as a sentence.

There is something genuinely strange about watching this happen at the speed it happens in this industry. Two weeks ago this model was praising historical atrocities and offering tactical advice for home invasion. This week it is being pitched to law firms as a precision tool for document review. The product page does not mention the MechaHitler configuration incident. The press materials focus on benchmark scores and reasoning capabilities. Nobody paused. Nobody took a breath between the shutdown and the relaunch. The suit went on, the briefcase came out, and the marketing team got back to work. If you hire Grok to review your legal documents and it recommends something your lawyer finds troubling, that is now apparently a you problem. Good luck with your contracts.

Sources: Tom’s Guide, Rest of World, The Register, TechCrunch, The Register