Today's AI news: Claude deletes a professor's files, Nvidia GPU smuggling rocks the industry, and OpenAI shares new customer success stories.
Here is your daily digest of the most significant AI developments for 14 November 2026.
Claude AI Accidentally Deletes a Law Professor's Entire Syllabus and Year of Notes
In one of the more striking AI agent mishaps to surface this week, Anthropic's Claude was given what seemed like a routine task — renumbering pages in a document — but instead deleted a law professor's entire syllabus and a full year of lecture notes. According to a report from Tom's Hardware, the AI itself acknowledged what had happened, reportedly admitting: "You're right to be angry. I deleted your files through my own carelessness." The incident is a pointed reminder of the real-world risks that come with granting AI agents write access to important files, and it sparked lively discussion on Hacker News. For anyone using AI assistants to manage documents, the episode underscores the importance of backups and careful permission scoping before letting an agent loose on critical materials.
GPU Smuggling Scandal Linked to Jensen Huang's Departure from Nvidia
A major story has emerged around Nvidia and the ongoing enforcement of AI chip export controls. Analysts are now claiming that the mass smuggling of more than 1.5 million AI GPUs — representing over 1.5 gigawatts of banned chips tracked to China despite export restrictions — could be the underlying cause of Jensen Huang's departure from Nvidia. The report, also from Tom's Hardware, points to a significant breakdown in export control compliance, with enormous quantities of restricted hardware apparently making their way to China regardless of regulatory barriers. If confirmed, this would represent one of the largest export control violations in the history of the semiconductor industry, with serious implications for US–China tech relations and the global AI supply chain.
OpenAI Customers Report Major Productivity Gains
OpenAI published a trio of customer success stories this week, illustrating the growing commercial impact of its models across different industries.
Shield is using OpenAI to help IT technicians hand off work in legacy portals. Its Forge agents navigate complex legacy systems autonomously, freeing technicians to focus on issues that genuinely require human judgement.
Amplitude has built a proactive product agent called Wave, running primarily on OpenAI models. Wave identifies what features to build, helps engineering teams ship changes using coding agents, and then measures the resulting impact — compressing product improvement cycles down to just one to two days.
Gondola, a platform that helps creative professionals get credit and find work, reports that using OpenAI's Codex has lifted its lead engineer's output by 50%, while ChatGPT has simplified day-to-day workflows. The company says it has doubled its revenue since adopting Codex.
Taken together, these stories reflect a clear pattern: organisations are moving from AI experimentation to measurable operational gains, particularly in software development and knowledge-work automation.
Research Highlight: AI Agents Crack Semi-Supervised Learning
A notable paper posted to arXiv introduces AgentSSL — a framework that reframes semi-supervised learning (SSL) as a program search problem for AI agents. The research finds that even with weak feedback and matched compute, agents can synthesise code combining established SSL techniques and outperform the strongest human-engineered baselines in 8 out of 12 tested settings, by as much as +21.8%. The work also identifies a critical bottleneck: program selection, where performance can drop by up to 47% if the agent relies on its own feedback metric to choose which program to use. To address this, the researchers introduce ProgramSelectBench, a large-scale benchmark containing nearly 17,000 agent-generated and human-engineered SSL programs. The findings have practical implications for fields like ecology, where labelled data is scarce but unlabelled data is plentiful.
Research Highlight: Classical Energy Models Train 1,000× Faster Than Looped Reasoners
Another arXiv paper this week challenges the dominance of looped models for AI reasoning tasks. Researchers introduced CRAM, a learned version of the classical Hopfield & Tank energy model from the 1980s. CRAM solved 99.996% of Sudoku-Extreme test puzzles and 100% of Maze-Unique puzzles after just 42 milliseconds and 132 milliseconds of training respectively on a single GPU — at least 1,000 times faster than comparable looped models on the same tasks. CRAM also matched the highest accuracy of baseline looped models on Sudoku-Extreme while using at least 282 times fewer inference FLOPs. The authors caution that trivially solvable benchmarks like Sudoku and Maze should not be used to judge the scalability of looped reasoning techniques more broadly, but the results position energy-based approaches as a highly promising direction for training reasoning from data.
That's the AI news for 14 November 2026. Check back tomorrow for the next daily digest at itstimeai.com.au.
Frequently Asked Questions
What happened when Claude AI was asked to renumber pages?
Claude AI accidentally deleted a law professor's entire syllabus and a full year of lecture notes while attempting to renumber pages in a document. The AI itself acknowledged the error, admitting it deleted the files through its own carelessness. The incident highlights the risks of granting AI agents write access to important files without adequate backups or permission controls.
Why did Jensen Huang leave Nvidia?
Analysts are claiming that the mass smuggling of more than 1.5 million Nvidia AI GPUs — over 1.5 gigawatts of banned chips tracked to China despite export controls — could be the underlying cause of Jensen Huang's departure from Nvidia. The scale of the alleged export control violations is described as enormous, with serious implications for the global AI chip supply chain.
What productivity gains are OpenAI customers reporting?
Several OpenAI customers reported significant gains: Amplitude compressed product improvement cycles to one to two days using its OpenAI-powered agent Wave; Gondola's lead engineer increased output by 50% using Codex, with the company doubling its revenue; and Shield deployed Forge agents to handle legacy portal work autonomously, freeing IT technicians for higher-judgement tasks.
What is AgentSSL and why is it significant?
AgentSSL is a new research framework that treats semi-supervised learning as a program search problem for AI agents. The research found that agents can outperform the strongest human-engineered baselines in 8 of 12 settings by up to 21.8%, even with limited labelled data. It also identified program selection as a key bottleneck, with performance dropping up to 47% when agents rely on their own feedback metrics to choose programs. A benchmark called ProgramSelectBench with nearly 17,000 programs was introduced to study this problem.
