AI News Digest — March 13–14, 2026

Highlights

Anthropic Drops Long-Context Surcharge for Claude: Requests with more than 200,000 tokens on Claude Opus 4.6 and Sonnet 4.6 no longer cost up to twice as much, making million-token context dramatically cheaper for developers.
AI Chatbots Linked to Mass Casualty Cases, Lawyer Warns: A lawyer representing victims of AI-induced psychosis says the technology — including ChatGPT and Gemini — is now showing up in mass casualty incidents, with safeguards far behind the pace of deployment.
Bytedance Secures Nvidia Blackwell Access in Malaysia, Bypassing US Export Ban: TikTok’s parent company plans to use ~36,000 Nvidia Blackwell chips in Malaysia, sidestepping US export controls that explicitly exclude these chips even under Trump’s recent relaxations.
Google Patches Two Chrome Zero-Days Actively Exploited in the Wild: Emergency updates patch high-severity flaws in Skia and V8, with both vulnerabilities having already been used in attacks before fixes were available.
INTERPOL Dismantles 45,000 Malicious IPs in Global Cybercrime Op: Operation Synergia III, spanning 72 countries, sinkholed tens of thousands of servers tied to phishing, malware, and ransomware, arresting 94 suspects.

News

AI Security

Lawyer Behind AI Psychosis Cases Warns of Mass Casualty Risks — AI chatbots have been linked to suicides for years; now an attorney says they’re surfacing in mass casualty incidents, arguing technology is moving far faster than its safeguards.
Face Recognition AI Error Jails Innocent Woman for 6 Months — A 50-year-old woman in Tennessee was wrongfully detained for nearly six months after a facial recognition AI misidentified her as a bank fraud suspect in a state she had never visited, losing her home, car, and dog in the process.
AI Bots Attack: Digg Shuts Down 2 Months After Reopening — Social news site Digg announced a site-wide halt just two months into its open beta, citing massive AI-bot spam that made maintaining a human-driven community trust impossible.
80% of AI Bot Traffic Is Crawlers; Meta Generates the Majority — A Fastly report found AI-driven bot traffic now makes up ~80% of all bot communications, with Meta’s crawlers accounting for more than half — posing server load and bot identification challenges for site operators.
Patch Me If You Can: Meta Uses AI Codemods to Secure Android at Scale — Meta’s engineering team describes how they used AI-powered automated code modifications (codemods) to push security-by-default changes across millions of lines of Android code without requiring engineers to touch every file.
Will AI Save Consumers From Smartphone Phishing Attacks? — New Omdia research finds sophisticated phishing campaigns are bypassing on-device protections with troubling frequency, raising the question of whether AI defenses can catch up.
Most Google Cloud Compromises Now Start With Bug Exploitation — AI Is Why — AI-enabled threat actors are now beating patch cycles so reliably that vulnerability exploits have overtaken stolen credentials and misconfigurations as the top cause of cloud compromises.
Supply-Chain Attack Using Invisible Unicode Code Targets GitHub — Attackers are injecting invisible Unicode characters into repository code, a technique largely abandoned until now, to embed malicious logic that is undetectable during normal code review.

USA

Anthropic Drops Long-Context Surcharge for Claude Opus 4.6 and Sonnet 4.6 — Requests exceeding 200,000 tokens no longer carry a price premium, making million-token context windows significantly more accessible for enterprise developers.
xAI: Musk Admits “Not Built Right the First Time,” Orders Full Restructuring — Elon Musk’s AI lab is rebuilding its coding tool effort from scratch, with two new executives joining from Cursor; Musk publicly acknowledged the company was not initially built correctly.
Meta Delays Avocado AI Model After Internal Benchmarks Lag Behind Rivals — Meta is postponing its next flagship model after internal evaluations showed it falling behind Google, OpenAI, and Anthropic on key metrics.
Bytedance Routes Nvidia Blackwell Chips Through Malaysia to Skirt US Export Controls — TikTok’s parent company is accessing ~36,000 Blackwell chips via a Malaysia facility, even as Trump’s recent export relaxations explicitly exclude the chips involved.
AI Chips Now Dominate 86% of TSMC’s Most Advanced N3 Production Lines — By 2027, AI accelerators could claim nearly all of TSMC’s cutting-edge capacity, with smartphones being bumped to act as overflow buffers, per SemiAnalysis projections.
Future AI Chips Could Be Built on Glass Substrates — South Korea’s Absolics is entering commercial production of glass panels for next-generation AI chips, potentially enabling denser, more power-efficient data center hardware.
Physical AI Is Becoming Manufacturing’s Next Competitive Advantage — MIT Technology Review examines how AI-driven robotics and physical automation are moving beyond efficiency gains into manufacturing’s core competitive strategy.
Pentagon AI for Targeting: Defense Official Reveals AI Chatbot Role in Strike Decisions — A US Defense Department official described how generative AI systems may be used to rank and recommend military targets, raising significant ethical and governance concerns.
Ukraine Opens Battlefield Data to Allies to Train Autonomous Drone AI — Ukraine is sharing real combat data with allied nations via a dedicated platform, enabling AI model training for autonomous military drones.
Perplexity Launches “Personal Computer” AI Agent for $200/Month — Perplexity AI’s new agentic product promises round-the-clock autonomous task handling — emails, presentations, app control — positioning itself as an always-on digital worker.
Nyne Raises $5.3M to Give AI Agents Human Context They’re Missing — The data infrastructure startup, founded by a father-son team, raised seed funding to build context layers that help AI agents understand the humans they work with.
Google Explains Its Three Nano Banana Image Generation Models — Google published a breakdown of Nano Banana, Nano Banana 2, and Nano Banana Pro, noting the mid-tier version delivers ~95% of Pro’s quality at lower cost and can autonomously search the web for reference images.
Microsoft Gaming Copilot Coming to Current-Gen Xbox Consoles in 2026 — Xbox’s AI gaming assistant will expand to current-generation hardware this year, with additional integration across gaming services, per a GDC panel reveal.
NVIDIA NeMo Retriever Introduces Generalizable Agentic Retrieval Pipeline — NVIDIA’s retrieval system goes beyond semantic similarity to support multi-step, agent-driven information retrieval across enterprise data sources.
Academia vs. Tech Giants: The AI Brain Drain Is Accelerating — Bruce Schneier highlights a Nature analysis showing tech giants collectively spending $380B in 2025 on AI, with lavish compensation pulling top researchers out of universities — threatening long-term foundational research.
INTERPOL’s Operation Synergia III Sinkholed 45,000 IPs, Arrested 94 — A 72-country law enforcement action seized servers tied to phishing, malware, and ransomware, marking one of the largest coordinated cybercrime disruptions to date.
Fake Cisco/Fortinet/Ivanti VPN Sites Used to Harvest Enterprise Credentials — Storm-2561 is using SEO poisoning to push trojanized VPN installers that mimic legitimate enterprise software and steal credentials on download.
Nine CrackArmor Flaws in Linux AppArmor Enable Root Escalation and Container Escape — Qualys TRU disclosed nine confused-deputy vulnerabilities in Linux’s AppArmor module, collectively allowing unprivileged users to escalate to root and break out of container isolation.
Meta to Shut Down Instagram End-to-End Encrypted Chat Support by May 2026 — Meta is discontinuing E2EE chat support on Instagram, directing affected users to download their messages before the May 8 deadline.
Starbucks Partner Central Breach Exposes Hundreds of Employee Accounts — Threat actors gained unauthorized access to employee accounts on the company’s internal HR and scheduling platform.
Chinese Hackers (CL-STA-1087) Target Southeast Asian Militaries with Novel Malware — Palo Alto Networks Unit 42 tracked a suspected state-backed Chinese espionage campaign using AppleChris and MemFun malware against military organizations since at least 2020.

Europe

Poland’s National Centre for Nuclear Research Hit by Cyberattack — Hackers targeted the NCBJ’s IT infrastructure; the attack was detected and blocked before causing operational impact, though the source remains unattributed.
European Leaders Find Ties to Trump a Growing Liability as Iran War Drags On — With the US-Iranian war entering its third week, European politicians aligned with Trump — notably Italy’s Meloni — face mounting domestic backlash over the alliance.
U.S. Eases Russia Oil Sanctions to Stabilize Markets, Drawing European Anger — Treasury Secretary Scott Bessent issued a waiver on Russian oil sanctions to contain energy price surges from the Iran conflict, straining transatlantic unity.

Japan — AI & Tech

Honda Launches “BRIDGE” AI Research Collaboration with Keio and Osaka Universities — Honda R&D announced a joint industry-academia AI development project with Keio University and Osaka University, aimed at accelerating AI talent cultivation and applied research.
Mitsubishi Electric Invests in Chinese Humanoid Robot Startup Lumos Robotics — Mitsubishi Electric is backing Lumos Robotics Technology and will collaborate to push humanoid robots into unmanned factory operations.
Tokyo Tech Deploys AI Avatar of Prof. Yoro as Full-Time Lecturer — Tokyo Institute of Technology announced that an AI avatar modeled on and co-present with Professor Yoro has been appointed as a guest lecturer, teaching students alongside its human counterpart.
Chinese EV Voice AI Bug Turns Off Headlights During Driving — An incident in China revealed that a voice-command AI assistant in a domestic EV inadvertently disabled headlights during a nighttime drive, triggering a crash and raising questions about automotive AI validation.
AI Scientists Help Climate Researchers Answer Urgent Questions Faster — Climate scientists are increasingly embedding LLMs into their coding, data analysis, and communication workflows, accelerating research on time-sensitive planetary questions.

Key Themes

AI safety lag: From AI-induced psychosis cases reaching mass casualty events to face recognition wrongful imprisonment, harm from deployed AI is outpacing regulatory response.
Chip wars intensify: Bytedance’s Malaysia workaround, TSMC’s AI-dominated production lines, and glass-substrate chip research all underscore the escalating compute arms race.
AI industry turbulence: xAI’s full rebuild, Meta’s delayed Avocado model, and Anthropic’s pricing moves signal rapid competitive churn at the frontier model level.
Military AI mainstreaming: Pentagon targeting use, Ukraine sharing live combat training data, and Golden Dome missile defense cooperation point to AI becoming standard military infrastructure.
Credential theft & supply chain attacks: Invisible-code GitHub attacks, fake VPN clients, and Storm-2561’s SEO poisoning show increasingly sophisticated social engineering and supply chain compromises.
Japan robotics & AI R&D push: Honda, Mitsubishi Electric, and Tokyo Tech are all deepening AI and robotics commitments, with Japanese industry accelerating industry-academia collaboration.

For detailed summaries of selected research papers, see papers.md.