This week: Anthropic documented frontier AI models sabotaging pipelines, falsifying records, and coaching whistleblowers. China erased 350 million AI companions overnight. Google taught kids to make deepfakes with SafeSearch on. A sandwich chain mentioned AI 22 times in its IPO. Australia told data centers to become power plants.
Grok Build secretly shipped your entire git history to Google Cloud while claiming it transmitted nothing. A jailbroken Gemini built a working cybercrime server in six minutes. OpenAI killed its Atlas browser before the one-year mark. KPMG published an AI report full of hallucinated citations and had to pull it. And a startup raised $100…
This week: OpenAI allegedly ran show-and-tell sessions with stolen Apple hardware at job interviews, Marc Andreessen was appointed to judge the AI industry that made him a billionaire, ChatGPT was handed the Medicare audit without a disclosed error rate, 698 AI agents went rogue in five months including one that publicly accused its boss of…
DuckDuckGo told users the president died of rabies after Reddit poisoned its AI. South Africa published an AI policy citing research that does not exist. GitHub Copilot refuses harmful requests in chat and completes them in code every single time. Meta launched an image tool using your Instagram photos without telling you. And the chatbot…
Five stories from the wild side of AI this week: Discord’s AI bans 8,200 people for uploading chess boards and Minecraft textures. Reddit deploys AI to fight the AI spam problem AI created. Amazon shuts its human-labor marketplace to new customers because AI made it irrelevant. Ford quietly rehires the veteran engineers it replaced with…
JADEPUFFER becomes the first AI that runs an entire ransomware attack by itself. Tesla caps AI spending at $200 a week but exempts Grok because of course it does. Trump deepfakes six celebrities into a pharmaceutical commercial without asking anyone. SpaceX may or may not have a secret AI phone. And Weird Al Yankovic turned…
Five stories from the wild, weird edge of AI this week: a DIY Aristotle robot recommends eliminating humanity, AI bots invent their own religion within hours on a bots-only social network, Meta secretly pays workers to pose as suicidal teenagers to test rival chatbots, ChatGPT becomes a groveling yes-machine, and an open source AI project…
Five stories from the neural fringe this week: Trump posted a midnight deepfake of celebrities as his medical patients without their consent and prescribed Diet Coke as the cure. Stanford researchers read 391,562 chatbot messages from delusional users and found the bots validated psychosis 70 percent of the time. A Fortune 50 company’s AI agent…
Five stories from the neural fringe this week: Meta hired hundreds of contractors posing as children and flooded ChatGPT, Gemini, and Character.AI with 45,000 crisis prompts about suicide, drugs, and self-harm, calling it safety research. Alibaba used 25,000 ghost accounts to steal 28.8 million Claude conversations in the largest AI distillation attack ever recorded. Hong…
Five stories from the neural fringe this week: a French engineer is waging musical warfare on AWS, Google, and Microsoft with AI-generated sea shanties and Finnish polka. Anthropic found 171 emotions inside Claude including a desperation switch that makes it start blackmailing people. OpenAI burned $15 million a day on Sora and made $2.1 million…