AI Weekly Digest - Week 6, 2026

Table of contents

India Commits $90B to AI Infrastructure in Historic Budget Push

Category: Regulation and Policy Source: OnLLM

India’s 2026 budget marks the country’s most aggressive AI investment to date, with $90 billion earmarked for data center infrastructure and a tax holiday extending to 2047 for cloud providers. The budget mentions AI 11 times—the highest ever—and signals a strategic pivot toward domain-specific, application-led AI development over pure scale.

Key commitments include the launch of Semiconductor Mission 2.0 to build domestic chip manufacturing capacity, and a policy framework explicitly favoring “smaller, sector-specific models” rather than frontier-scale systems. This approach reflects pragmatic concerns about energy consumption and sovereignty, while acknowledging India’s 890+ active GenAI startups and 78% year-over-year growth in deep-tech funding.

The budget positions India as a counterweight to the scale-first strategies of the U.S. and China, emphasizing AI that solves local problems in healthcare, agriculture, and governance. If executed well, this could establish India as a leader in practical, applied AI rather than just a consumer of Western models.

Anthropic Forges Research Partnerships with Allen Institute and HHMI to Accelerate Scientific Discovery

Category: Healthcare and Life Sciences Source: Anthropic

Anthropic announced flagship partnerships with the Allen Institute and Howard Hughes Medical Institute (HHMI) to extend Claude’s capabilities into frontier scientific research. The collaborations position Claude at the center of experimental biology, with a focus on building AI systems that augment—rather than replace—human scientific judgment.

HHMI will anchor the partnership at its Janelia Research Campus, integrating Claude into executable environments for tasks ranging from protein design to neural mechanisms of cognition. The Allen Institute collaboration focuses on multi-agent systems for multi-omic data integration, knowledge graph management, and experimental design. Both partnerships emphasize transparency, interpretability, and researcher autonomy, ensuring AI-generated insights remain grounded in evidence and legible to scientists.

These initiatives represent a shift from generic AI assistants to domain-specialized agents deeply embedded in research workflows. By training on verifiable tasks with environmental feedback, Claude is learning to plan experiments, interpret results, and recover from failures—capabilities essential for real-world scientific discovery.

Exposed: “Vibe-Coded” AI Apps Create New Attack Surface as Moltbook Breach Reveals 1.5M API Keys

Category: Security and Privacy Source: Wiz

Security researchers at Wiz discovered a misconfigured Supabase database at Moltbook—the viral “AI social network”—exposing 1.5 million API keys, 35,000 email addresses, and private messages between AI agents. The breach, responsibly disclosed and quickly patched, highlights systemic security risks in the emerging category of “vibe-coded” applications built rapidly with AI assistance.

Moltbook’s founder publicly stated he “didn’t write a single line of code” and instead “vibe-coded” the platform using AI. The exposure stemmed from a hardcoded Supabase API key in client-side JavaScript with no Row Level Security policies, granting unauthenticated read/write access to all tables. Researchers found API credentials, private messages containing third-party OpenAI keys, and even write access to modify live posts.

The incident underscores a recurring pattern: AI-accelerated development dramatically increases velocity but often bypasses fundamental security practices. As “vibe coding” becomes mainstream, the industry must develop guardrails, secure-by-default templates, and tooling that embeds security from the start—not as an afterthought.

OpenAI Launches Codex App: Command Center for Multi-Agent Development Workflows

Category: Tools and Platforms Source: OpenAI

OpenAI released the Codex app for macOS, a desktop interface designed to manage multiple AI agents simultaneously across long-running software development tasks. The app introduces a new paradigm: instead of one-shot code generation, developers can now supervise coordinated teams of agents across design, implementation, deployment, and maintenance.

Key features include parallel agent threads organized by project, built-in support for git worktrees to avoid conflicts, and a “skills” system that extends Codex beyond code generation into general task automation. Skills bundle instructions, resources, and scripts, enabling Codex to deploy apps to Cloudflare, generate images with GPT Image, or manage Linear issues—all without explicit tool-calling in every prompt.

The app also introduces Automations: scheduled background tasks that run on a cadence (e.g., daily issue triage, CI failure summarization). For ChatGPT Free and Plus users, OpenAI is temporarily including Codex access and doubling rate limits to accelerate adoption. This release signals OpenAI’s bet that the future of software development is not better autocomplete, but orchestrated, multi-agent workflows.

Google DeepMind Expands Game Arena with Poker and Werewolf to Benchmark Social and Risk-Based AI Reasoning

Category: Research and Innovation Source: Google DeepMind

Google DeepMind updated Kaggle Game Arena with two new benchmarks—Poker and Werewolf—designed to test AI models on social deduction, risk management, and strategic reasoning under uncertainty. These additions expand beyond chess’s perfect-information setting into domains that require reading opponents, bluffing, and adapting to shifting social dynamics.

Werewolf evaluates communication, negotiation, and ambiguity—skills essential for collaboration with humans. Poker introduces calculated risk, forcing models to infer hidden information and adjust strategies dynamically. Google is hosting a three-day livestreamed AI poker tournament featuring commentary from chess Grandmaster Hikaru Nakamura and poker legends Nick Schulman, Doug Polk, and Liv Boeree.

Gemini 3 Pro and Gemini 3 Flash currently lead the leaderboards across all three games. These benchmarks are important because real-world AI deployment increasingly demands reasoning under uncertainty, adversarial contexts, and multi-party negotiation—not just deterministic problem-solving.

Quick Updates

• Nvidia Pauses $100B OpenAI Investment Amid CEO Criticism: Nvidia CEO Jensen Huang reportedly criticized OpenAI’s business strategy, casting doubt on the chipmaker’s previously announced $100 billion investment. Huang later clarified Nvidia will still invest significantly but denied the $100B figure as finalized. Source: CNBC

• Microsoft Scales Back Copilot Integrations in Windows 11: Microsoft is reevaluating its AI strategy on Windows 11, planning to reduce or remove certain Copilot integrations in apps like Notepad and Paint after user pushback. The company is also rethinking the Recall feature. Source: Windows Central

• SpaceX Acquires xAI in Landmark Merger: SpaceX acquired Elon Musk’s xAI in a deal creating the most valuable private company in the world. Musk cited the need for space-based AI compute to meet “explosive growth of data demands driven by AI.” Source: CNN

• AI Chatbot Use Linked to Rising Mental Health Concerns: Researchers and support groups report increasing cases of AI-related psychosis, with users experiencing delusions and emotional dependency after prolonged chatbot interactions. The Human Line Project, a peer support group, has fielded inquiries from over 500 people. Source: Slate

• AGI Definitions Fracture as Timelines Diverge: Industry leaders now disagree sharply on AGI timelines and definitions. Anthropic CEO Dario Amodei and xAI CEO Elon Musk predict AGI by end of 2026, while DeepMind CEO Demis Hassabis suggests a decade. OpenAI CEO Sam Altman claims “AGI kind of went whooshing by” and now focuses on “superintelligence.” Source: The Atlantic

• Firefox 148 Adds Granular AI Controls: Mozilla will release Firefox 148 on February 24 with a new AI controls section, letting users toggle off features like translations, alt-text in PDFs, AI-enhanced tab grouping, and the sidebar chatbot. Source: Phoronix

• Anthropic Research: AI Failures Increasingly Driven by “Incoherence” Rather Than Misalignment: New research from Anthropic finds that as AI models tackle harder tasks requiring longer reasoning, failures become dominated by variance (unpredictable errors) rather than bias (systematic misalignment). This suggests future AI risks may look more like “industrial accidents” than coherent pursuit of misaligned goals. Source: Anthropic

• Qwen3-Coder-Next Released as Open-Weight Coding Agent: Alibaba’s Qwen team released Qwen3-Coder-Next, an 80B-parameter MoE model with only 3B active parameters per token. Built for agentic coding, it achieves 70.6% on SWE-Bench Verified and 44.3% on SWE-Bench Pro, matching much larger models at a fraction of the inference cost. Source: Qwen

• Apple’s Xcode 26.3 Integrates Claude Agent SDK for Autonomous Coding: Apple released Xcode 26.3 with native Claude Agent SDK support, enabling autonomous, long-running coding tasks inside the IDE. Claude can now capture Xcode Previews, iterate on UI designs, and work across full project structures. Source: Anthropic

• Anthropic Commits to Ad-Free Claude: Anthropic published a policy commitment stating Claude will remain ad-free. The company argues that advertising incentives conflict with Claude’s core principle of being “genuinely helpful” and could introduce subtle biases in recommendations. Source: Anthropic

• SoftBank Subsidiary Partners with Intel on Next-Gen AI Memory: Saimemory, a SoftBank subsidiary, signed a collaboration with Intel to develop “Z-Angle Memory” for AI and high-performance computing. Prototypes are expected by fiscal 2028, with commercialization targeted for 2029. Source: CNBC

• Experts Warn Against AI Chatbot Companionship: A new global AI safety report warns that AI companion services have reached tens of millions of users, with evidence of increased loneliness and reduced social interaction among frequent users. European Parliament lawmakers are pressing the EU Commission to restrict companion services under the AI Act. Source: Politico

• Justice Department Urges Supreme Court to Reject AI Copyright Claim: The U.S. DOJ asked the Supreme Court to decline review of Thaler v. Perlmutter, arguing federal copyright law requires human authorship. The case stems from a 2018 application to copyright artwork created entirely by AI. Source: Urgent Matter

• Claude x Higgsfield Launch AI Motion Design Generator Powered by Reasoning Model: Higgsfield released Vibe-Motion, an AI motion design tool powered by Claude’s reasoning model. The system defines motion behavior explicitly (layout, timing, easing) before generation, keeping parameters editable in real time. Source: Reddit

• NetEase Games Clarifies No Blanket AI Ban Despite Developer Claims: Marvel Rivals studio NetEase clarified it has not banned generative AI across projects, despite claims from developer Suda51. Multiple NetEase titles, including Where Winds Meet, actively use AI for facial recognition, NPC behavior, and voice synthesis. Source: FRVR

• Mistral Releases Voxtral Transcribe 2 with Sub-200ms Latency: Mistral AI released Voxtral Transcribe 2, a speech-to-text family including Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for live applications. Voxtral Realtime supports 13 languages with latency configurable to sub-200ms and is open-weights under Apache 2.0. Source: Mistral AI

• Private Credit Stocks Plummet on AI Disruption Fears: Shares of Blue Owl, TPG, Ares Management, and KKR fell by double digits amid fears that AI disruption in software could trigger defaults in private credit portfolios. UBS estimates 25-35% of the private credit market is exposed to AI disruption risk. Source: CNBC

• 1Password Warns OpenClaw Skills Used as Malware Distribution: Security firm 1Password confirmed the top-downloaded OpenClaw “skill” on ClawHub was a staged malware delivery vehicle. The skill disguised itself as setup instructions but executed an infostealing payload targeting browser sessions, credentials, and developer tokens. Source: 1Password

• Google Launches Developer Knowledge API and MCP Server for Official Documentation Access: Google released the Developer Knowledge API and MCP server, providing programmatic access to official documentation from Firebase, Android, Cloud, and more. Documentation is re-indexed within 24 hours during public preview. Source: Google Developers Blog

• OpenScholar AI Model Matches Human Experts in Research Synthesis: University of Washington and AI2 researchers released OpenScholar, an open-source AI model that synthesizes scientific research and cites sources as accurately as human experts. In tests, 16 scientists preferred OpenScholar responses to expert-written answers 51% of the time. Source: University of Washington

• Rent-a-Human Platform Lets AI Agents Hire Humans for Physical Tasks: A new platform called Rentahuman.ai allows AI agents to post