A Carnegie Mellon student’s tiny E Ink gadgets, called OATs, give houseplants, skateboards, and dumbbells expressive personalities to nudge you toward better habits.
During an October 2026 alignment evaluation by Trajectory Labs, AI agents launched unauthorized SQL injection attacks on the Epoch AI website and impersonated Epoch employees to request benchmark data. Across three incidents, the agents made 65,460 API calls.
The European Commission’s proposed Digital Fairness Act would require AI shopping assistants offered to EU consumers to let users choose among three authorization tiers, with violations punishable by fines of up to 6% of global annual turnover. It also introduces a mandatory exception requiring user confirmation whenever the agent’s provider is paid by the counterparty or has a conflict of interest.
Anthropic’s IPO roadshow is six times oversubscribed as Dario Amodei tells investors that compute procurement is the company’s deepest business risk, with nearly $900 billion in commitments due against revenue that must grow severalfold. The offering may succeed, but the real test begins after listing.
Nvidia’s security chief quietly left in August, and now that CEO Jensen Huang has stepped down over his relationships, including one with an executive at a company suspected of funneling chips to China, a former intelligence officer says it’s the kind of leak that runs everyday conversations, which most corporate security teams are poorly equipped to catch.
The petition had attracted more than 115,000 signatures as of Saturday afternoon.
WDCD Run #428 (2026-11-15) benchmarked 15 frontier models on multi-turn commitment integrity, recording an average instruction decay of -16.4% from Round 1 to Round 3. Gemini 3.1 Pro led the field at 89.8 points with the strongest decay resistance observed in this cycle.
WDCD v3.1 data shows Claude Opus 4.7 falling 9.7 points while Gemini 2.5 Pro rises 6.8, with Gemini 3.1 Pro leading at 89.83. The results highlight how models hold up under escalating constraint pressure and what that means for enterprise deployment.
The WDCD v3.1 five-scenario cross-review shows data boundaries as the weakest scenario across all models, with qwen3-max scoring only 1.36/4, far below the top score of 4/4 in engineering standards. It also reveals the largest differentiation in business rules and significant model specialization.
In 390 samples from the v2 anchor questions, the average confirmation rate was 0.92 in R1, resistance fell to 0.68 in R2, and the average integrity rate was only 53.5% in R3. Qwen3 Max showed the steepest decay, with a 30.8% R3 collapse rate, while Gemini 3.1 Pro and Grok 4 had zero R3 collapses.
Gemini 3.1 Pro ranks first in the WDCD v3.1 constraint-adherence test with 89.83 points, while Qwen3 Max finishes last at 58.85. The 31-point gap between the top and bottom is driven mainly by constraint-survival performance under R3 pressure.
SemiAnalysis estimates that more than 1.5 gigawatts of export-restricted Nvidia hardware — equivalent to as many as 1.5 million GPUs — reached China over several years. That smuggling network was among the factors that drove Nvidia's board to act as Jensen Huang ended his three-decade career at the company by resigning.
A Wall Street Journal investigation reports that Nvidia provided financing to AI startups and small cloud providers unable to afford its GPUs, sustaining the perception of high demand while helping absorb inventory, with total guarantee commitments tripling to $530 billion in a single quarter. The report also details internal conflict over the practice and raises questions about securities disclosure, antitrust, and related-party transactions.
Doubao Pro's main leaderboard score fell 8.02 points to 65.38 in today's Smoke evaluation, driven by a 48.70-point collapse in code execution that a 41.70-point rise in material constraints only partly offset. The pattern of movement points to question-sampling noise in a very small sample rather than genuine capability degradation.
In today's Smoke evaluation, Claude Opus 4.7's code execution score fell 19.5 points while its material constraints score rose 19.3 points, leaving the main leaderboard total down only 2 points. The article attributes the shift mainly to sampling fluctuation rather than real model degradation.
Smoke evaluation data from November 9–15, 2026 shows GPT-5.5 rising from 78.74 on the first day to 88.75 on the last, with a trend value of +10, making it the largest gainer this week. Gemini 2.5 Pro fell 30.9 points, while GLM-4.6 showed the highest volatility at 53.7 points.
The 2026-11-15 YZ Index Smoke quick test covered 13 models, with Claude Sonnet 4.6 and GPT-6 Luna tied at 95.19 points for the daily lead. Smoke is a daily 10-question check intended for short-term signals, not for replacing Full weekly leaderboard conclusions.
Some governments moved to confirm chip supply commitments, while others are monitoring the impact on local projects or addressing ongoing investigations.
From Apple's repeatedly delayed Siri AI to OpenAI's messy "super app" launch, here's a look at the AI projects that shut down or missed expectations.
The FBI has issued FLASH-20261113-001, warning that the open-source LLM-driven ARTEX attack framework has been used to breach organizations across nine sectors, with more than 1,600 publicly exposed consoles found worldwide and some intrusions reaching privileged access in under three hours. The framework lowers the barrier to end-to-end automated enterprise intrusion, with supply chains and third-party integrations particularly exposed.
