A whistleblower trapped inside a "pig butchering" scam compound gave WIRED a vast trove of its internal materials--including 4,200 pages of messages that lay out its operations in unprecedented detail. Just before 8am one day last April, an office manager who went by the name Amani sent out a motivational message to his colleagues and subordinates. "Every day brings a new opportunity--a chance to connect, to inspire, and to make a difference," he wrote in his 500-word post to an office-wide WhatsApp group. "Talk to that next customer like you're bringing them something valuable--because you are." He and his underlings worked inside a " pig butchering " compound, a criminal operation built to carry out scams --promising romance and riches from crypto investments--that often defraud victims out of hundreds of thousands or even millions of dollars at a time. The workers Amani was addressing were eight hours into their 15-hour night shift in a high-rise building in the Golden Triangle special economic zone in Northern Laos. Like their marks, most of them were victims, too: forced laborers trapped in the compound, held in debt bondage with no passports. They struggled to meet scam revenue quotas to avoid fines that deepened their debt.

Roblox's AI-Powered Age Verification Is a Complete Mess Kids are being identified as adults--and vice versa--on Roblox, while age-verified accounts are already being sold online. Just days after launching, Roblox's much-hyped AI-powered age verification system is a complete mess. Roblox's face scanning system, which estimates peoples' ages before they can access the platform's chat functions, rolled out in the US and other countries around the world last week, after initially launching in a few locations in December. Roblox says it is implementing the system to allow users to safely chat with users of similar ages. But players are already in revolt because they can no longer chat to their friends, developers are demanding Roblox roll back the update, and crucially, experts say that not only is the AI mis-aging young players as adults and vice versa, the system does little to help address the problem it was designed to tackle: the flood of predators using the platform to groom young children.

Modern work relies on an assortment of digital collaboration tools, yet routine processes continue to suffer from human error and delay. To address this gap, this dissertation extends TheAgentCompany with a finance-focused environment and investigates whether a general purpose LLM agent can complete representative wealth-management tasks both accurately and economically. This study introduces synthetic domain data, enriches colleague simulations, and prototypes an automatic task-generation pipeline. The study aims to create and assess an evaluation set that can meaningfully measure an agent's fitness for assistant-level wealth management work. We construct a benchmark of 12 task-pairs for wealth management assistants spanning retrieval, analysis, and synthesis/communication, with explicit acceptance criteria and deterministic graders. We seeded a set of new finance-specific data and introduced a high vs. low-autonomy variant of every task. The paper concluded that agents are limited less by mathematical reasoning and more so by end-to-end workflow reliability, and meaningfully affected by autonomy level, and that incorrect evaluation of models have hindered benchmarking.

Practicing good "operations security" is essential to staying safe online. Here's a complete guide for teenagers (and anyone else) who wants to button up their digital lives. Teenagers have always been formidable hackers. In fact, in recent years, some of the most high-profile and brazen digital attacks around the world have been carried out by teens. And whether you've never given much thought to your digital privacy and security or you've started to rein in your data, you can use this guide to implement basic precautions and keep operations security in mind. In other words, this guide contains advice and ideas to help you conceptualize how people can find out information about you from your digital activities--and start to minimize what's out there in ways you didn't intend. Some people are more private by nature, and others prioritize putting themselves out there. But even if you're a 24/7 streamer, you can still think about your operations security, commonly known as opsec. What can viewers see in your room while you're streaming?

Public LLMs such as the Llama 2-Chat underwent alignment training and were considered safe. Recently Qi et al. [2024] reported that even benign fine-tuning on seemingly safe datasets can give rise to unsafe behaviors in the models.

Large language model systems face significant security risks from maliciously crafted messages that aim to overwrite the system's original instructions or leak

The ability of large language models (LLMs) to mimic human-like intelligence has led to a surge in LLM-based autonomous agents. Though recent LLMs seem capable of planning and reasoning given user instructions, their effectiveness in applying these capabilities for autonomous task solving remains underexplored. This is especially true in enterprise settings, where automated agents hold the promise of a high impact.

Large language models (LLMs) have seen increasing popularity in enterprise applications where AI agents and humans engage in objective-driven interactions. However, these systems are difficult to evaluate: data may be complex and unlabeled; human annotation is often impractical at scale; custom metrics can monitor for specific errors, but not previously-undetected ones; and LLM judges can produce unreliable results. We introduce the first set of unsupervised metrics for objective-driven interactions, leveraging statistical properties of unlabeled interaction data and using fine-tuned LLMs to adapt to distributional shifts. We develop metrics for labeling user goals, measuring goal completion, and quantifying LLM uncertainty without grounding evaluations in human-generated ideal responses. Our approach is validated on open-domain and task-specific interaction data.

Dick Cheney dead: Vice President who served with George W. Bush and took leading role in'war on terror' dies at 84 Why Tuesday's races aren't as close as you think: White House analyst CRAIG KESHISHIAN reveals what the polls always miss How Charles reacted when'difficult' William asked if he could do fewer engagements: Biographer ROBERT JOBSON reveals the'tension', secrets of Kate's family life - and how couple are making Prince George'strong' 'She used it to freshen up... it killed her': My wife died of cancer at 63. She never smoked or drank. Taylor Swift enjoys girls' night with squad member Gigi Hadid in NYC after Travis Kelce's ex took swipe Warren Buffett's $6billion stock exit is his loudest warning yet Father reveals'radical faith' spiral of American son killed in hail of arrows by reclusive tribe as new believers consider following him to isolated island My ex is the international fugitive Democrat who's fled to Europe with our nine-year-old son. Here are the disgraceful secrets she'd hate the world to know Woke Gen-Z's revenge: Poll reveals staggering number of under-30s who back Mamdani while their parents are terrified of a return to 1980s New York Coca-Cola faces huge backlash for using AI in its'Holidays are Coming' Christmas advert for the second year in a row - as one viewer calls it the'best ad I've ever seen for Pepsi' Explosive war sparked by Hollywood agent's nepo-baby son dismembering his wife Shohei Ohtani makes rare speech in English as Dodgers star's wife soaks up the spotlight at World Series parade It's the trendy new diagnosis for everything from fatigue to brain fog... but here's the truth about your gut problem - and how to fix it: DR EMILY LEEMING Dallas Cowboys agree huge trade for rival's defensive captain just hours before the NFL deadline Republicans join joyous liberals in dumping on Dick Cheney's death: 'Hell is hot!' AMANDA PLATELL: Fergie's delusions have reached a new low. I can't believe Beatrice and Eugenie are egging her on.

Large Language Models (LLMs) often expend significant computational resources generating boilerplate responses, such as refusals, simple acknowledgements and casual greetings, which adds unnecessary cost and latency. To address this inefficiency, we propose a simple yet highly effective method for detecting such responses after only a single generation step. We demonstrate that the log-probability distribution of the first generated token serves as a powerful signal for classifying the nature of the entire subsequent response. Our experiments, conducted across a diverse range of small, large, and reasoning-specialized models, show that the first-token log-probability vectors form distinctly separable clusters for different response types. Using a lightweight k-NN classifier, we achieve high accuracy in predicting whether a response will be a substantive answer or a form of boilerplate response, including user-specified refusals. The primary implication is a practical, computationally trivial technique, optimizing LLM inference by enabling early termination or redirection to a smaller model, thereby yielding significant savings in computational cost. This work presents a direct path toward more efficient and sustainable LLM deployment.