Brown University's AI Cheating Incident: Grades Drop ~50% on In-Person Exams

At Brown University, a prestigious Ivy League school, suspected AI-assisted cheating has sparked a major debate. A professor who suspected students of using AI to cheat switched from online exams to in-person final exams — and grades dropped by roughly 50%.

The story received prominent coverage on Ars Technica, accompanied by the line "We cannot be deliberately stupid," and drew a large response on Hacker News with 68 points and 44 comments. It's a case that makes you think again about how the spread of AI tools is affecting education.

Anthropic Cuts Claude Code's System Prompt by 80%

Anthropic has cut the system prompt of its AI coding tool, Claude Code, by 80%. The company's Tariq noted that excessive prohibition-style instructions for recent models such as Claude Fable 5 had been constraining the models' creativity.

The reasoning is that trimming the prompt allows AI agents to exercise more flexible judgment — an approach drawing attention as a way to unlock capability by reducing constraints. Japan's ITmedia also covered it as a breaking news item.

AI Costs Baffle the C-Suite — The Complexity of Usage-Based Pricing

According to The Register, as AI services shift to "usage-based pricing," C-suite executives increasingly struggle to keep tabs on their bills. Unlike traditional subscription models, costs fluctuate with actual usage, which makes budgets hard to forecast.

As AI adoption gets serious, it's clear that building proper cost management mechanisms has become urgent.

Qwen3.5 122B Is "the Best" — Hands-On Impressions from the LocalLLaMA Community

On Reddit's LocalLLaMA community, a lively discussion is underway about Qwen3.5 122B. The poster reports having tested multiple models on a machine with 128GB of RAM: Qwen3.6 27B/33B and Gemma4 31B/26B ran into problems on complex tasks, and in the end Qwen3.5 122B delivered the most consistent results.

Notably, on a real-world task extracting data from 160 PowerPoint files, Qwen3.5 122B completed the job in about two hours while other models kept running into issues. The observation that 120B-class MoE models occupy the "sweet spot" balancing capability and speed is a useful guideline for running local LLMs.

Hijacking Defensive AI Agents Enables Remote Code Execution

A report released by the AI Now Institute identifies a vulnerability in which AI agents used for cybersecurity defense can be hijacked, leading to remote code execution (RCE). Dubbed "Friendly Fire," this exploit outline highlights the risks of deploying AI agents for security purposes.

As a new threat vector in which AI itself becomes the attack target, it underscores the need to rethink security design.

---