ARS
Signal’s encryption just passed an AI-powered mathematical checkup
The four-week project hints that AI could make formal verification far more practical.
U.S.–China Summit Establishes Direct AI Incident Channel; Shelved Military Crisis Agreement Becomes the Biggest Hidden Risk
The United States and China have established a dedicated bilateral channel for AI-related incidents and plan a special AI dialogue in November. But the agreement is vague and deliberately avoids military issues, leaving the highest-risk scenarios outside its coverage.
Anthropic Launches Office Suite on First Day of IPO Roadshow as Microsoft Shares Fall 2.2%
On the same day it kicked off its IPO roadshow, Anthropic unveiled a suite of workplace products—Claude Sheets, Whiteboards and Data—that take direct aim at Microsoft 365, sending Microsoft shares down 2.2%. The launch is less about replacing Excel than about positioning Claude as the entry point to the enterprise workflow.
WD
A 13-Year-Old Asked AI About Suicide on Her School Laptop. It Didn’t Get Flagged for Weeks
The young girl’s parents allege that she used Gemini via her school Chromebook to plan her suicide, but the school’s monitoring systems didn’t know until it was almost too late.
TC
Nvidia’s crisis has corporate governance experts reaching for the popcorn
The governance experts we talked to say nobody has seen anything like what's unfolding right now.
South Korea's DEEPX Completes Tape-Out of Samsung 2nm Chip: 100B-Parameter Inference at 5W, a Watt-Level Turning Point for Edge AI?
DEEPX has taped out its second-generation DX-M2 AI inference chip on Samsung's 2nm GAA process, targeting large language and vision-language models at roughly 5W. First silicon is due in March 2027, with customer sampling in Q2 and initial mass production in Q4, ahead of full global production in 2028.
Fei-Fei Li Takes Atlas to CoRL Robotics Conference: What Exactly Is AMD Betting On with $8.2 Billion?
At CoRL 2026, Fei-Fei Li presented World Labs' Atlas world model alongside AMD CEO Lisa Su, their first joint public appearance since AMD announced its $8.2 billion acquisition of World Labs. The article examines what Atlas actually delivers and why AMD considers the research team and its world-model roadmap worth the price.
Glasswing Seven-Month Audit: Sharp Gap Between Anthropic's Vulnerability-Scanning Accuracy Claims and Measured Results
A VulnCheck audit released on November 9, 2026 found that Anthropic Project Glasswing reported 37,966 findings over seven months, with only 217 marked as fixed and only 2 of 442 attributed CVEs confirmed as exploited in the wild. The audit highlights a substantial gap between claimed true-positive rates and actual remediation and verification outcomes.
ARS
Ring cameras go mobile as Amazon debuts AI-powered security drone and robot
Amazon's roaming Ring Robotics devices will debut in late 2026 and early 2027.
TC
All 7 Anthropic founders will each cash out about $250M in IPO
Each of the seven co-founders plans to sell exactly the same number of shares in the IPO, which would bring each of them about $250 million in cash at the midpoint of the price range while leaving the group in control of the company.
ARS
What happens to mathematics when AI becomes an oracle?
Mathematician Steven Strogatz discusses his new book, Big Math: The Hidden Codes That Run Our World.
TC
OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect
New York is preparing to enforce its sweeping AI safety law, requiring major developers like OpenAI and Anthropic to register, disclose safety practices, and report critical incidents or face fines.
TC
OpenAI acquires Odessia, travel startup created by Sonder founder
OpenAI has not yet said what its plans are for the future of travel. But Sonder founder Francis Davidson is a well-known figure in the travel industry.
ARS
Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI
Researcher finds 35,000 Grok Build apps wide open to anyone who looked for them.
Google Releases Gemini Robotics 2.5: Behind the 96% Factory Success Rate, the Real Challenge of the Remaining 4%
Google DeepMind has released Gemini Robotics 2.5, with Apptronik's Apollo 2 humanoid achieving a 96% success rate sorting auto parts on Mercedes-Benz's active production lines. The model makes major gains on dexterity tasks, but the remaining 4% failure rate and a partner-only rollout show how far it still is from a plug-and-play platform product.
Anthropic Officially Launches IPO Roadshow: 825 Million Shares Priced at $70–$78, PBC Structure Poses Governance Risks
Anthropic has launched the roadshow for its Class A common stock IPO, offering 825 million shares at $70–$78 each and targeting a fully diluted valuation of up to $1.4 trillion. Its public benefit corporation structure and multi-class voting arrangements raise unusual governance questions for public investors.
Research Lab
WDCD Run #422: Average Instruction Decay Hits -66.7% as Claude Sonnet 4.6 Leads 15-Model Field
WDCD Run #422 (2026-11-11) evaluated 15 large language models across three rounds of multi-turn commitment testing, recording an average instruction decay of -66.7% between Round 1 and Round 3. Claude Sonnet 4.6 took the top spot with 93.9 points, followed by Gemini 3.1 Pro and GPT-6 Sol.
WDCD Cycle: GPT-6 Astra Falls 20.3 Points as Six Models Decline and Only Sonnet Rises
In the latest WDCD v3.1 commitment-adherence test, only Claude Sonnet 4.6 improved among the 15 evaluated models, while the other six models all declined, with GPT-6 Astra dropping 20.3 points. The results point to differences in how stably models maintain parallel hard constraints under multi-round pressure.
Safety Compliance Collapses to a Low of 1.7; Claude-Sonnet Dominates with 4/4 in Four Scenarios
In the WDCD v3.1 constraint-adherence test, safety compliance is the lowest-scoring scenario at just 1.7/4, while claude-sonnet-4.6 reaches 4/4 in four scenarios. Model strengths remain highly scenario-dependent, with safety compliance and data boundary requiring additional guardrails in production.
After an 87% R1 Confirmation Rate, R3 Still Collapses Twice: WDCD Anchor Task Shows Two Models Failing All Rounds
WDCD v2 anchor tasks covering 150 samples show an average R1 confirmation rate of 0.87 and an average R2 resistance rate of 0.73, with R3 full collapses occurring only twice. GLM-4.6 and Qwen3 Max scored zero across all three rounds, failing at the initial parsing stage.