AI Digest.

Axios Supply Chain Attack Hits npm as Claude Code Source Leaks via Source Maps

A critical supply chain attack on axios dominated the conversation today, with the npm package's 100M+ weekly downloads potentially compromised by injected malware. Meanwhile, Anthropic accidentally published Claude Code's full source via a .map file in their npm registry, and Google Research revealed a 20x improvement to Shor's algorithm that could break ECDSA encryption by 2029.

OpenAI Codex Vulnerability Exposes GitHub Tokens as Claude Code Gets Computer Use and Cross-Platform Plugins

A critical security vulnerability in OpenAI Codex allowed attackers to exfiltrate GitHub tokens through command injection via branch names. Meanwhile, Claude Code had a massive feature day with computer use capabilities and an official Codex integration plugin, while Alibaba dropped Qwen3.5-Omni with native multimodal understanding across text, image, audio, and video.

Self-Evolving Agents Go Open Source as Video Generation Models Challenge Commercial APIs

Today's feed centers on the rapid maturation of open-source AI agents and video models, with a standout project called 724 Office demonstrating self-repairing, tool-creating agent architecture in just 8 files. The community is also deep in conversation about agent design patterns, from structured disagreement protocols to reinforcement learning optimization, while Claude Code updates and rate limit frustrations reveal the growing pains of AI-native development workflows.

ARC-AGI 3 Stumps AI at Under 1% While Agent Infrastructure Race Heats Up

Today's feed centers on the growing ecosystem around AI agents, from integration platforms and communication protocols to the workflows developers are building around them. ARC-AGI 3 launched as a humbling new benchmark where AI scores below 1%, and the Claude Code ecosystem continues to expand with new hooks and memory features.

Anthropic's Leaked "Mythos" Model Sparks Safety Panic as Google Ships Gemini 3.1 Flash Live

A leaked Anthropic blog post revealing two upcoming models, Mythos and Capybara, dominated today's discourse with heated debate over AI safety and autonomous capabilities. Google quietly shipped Gemini 3.1 Flash Live for real-time voice agents, while the tooling ecosystem continued to mature with Cline Kanban for multi-agent orchestration and new Claude integrations for everything from tax prep to Mac setup automation.

TurboQuant Hits llama.cpp in 36 Hours as Agent Tooling Explodes Across Browsers, CLIs, and Scheduled Jobs

The open-source quantization community delivered a stunning paper-to-production pipeline with Google's TurboQuant running on Apple Silicon in under two days. Meanwhile, the agent ecosystem saw a wave of new tooling for browser automation, scheduled cloud jobs, and developer testing. Local inference benchmarks continued to push boundaries with NVIDIA's Mamba architecture outpacing Alibaba's DeltaNet on consumer GPUs.

LiteLLM Supply Chain Attack Exposes 97M Downloads as Google's TurboQuant Promises 6x Memory Compression

A major supply chain attack on LiteLLM's PyPI package dominated the conversation, with the compromise only caught because the attacker's malware was buggy enough to crash machines. Meanwhile, Google's TurboQuant algorithm for KV cache compression drew widespread excitement, and Cloudflare's Dynamic Workers launched as a new primitive for AI agent sandboxing.

NVIDIA's Nemotron Cascade Takes on Qwen While Claude Code Ships /schedule for 24/7 Automation

The AI agent ecosystem is maturing fast, with new tools for game dev, coding workflows, and autonomous scheduling dominating the conversation. On the model front, NVIDIA's Nemotron Cascade and on-device Qwen optimizations signal a shift toward efficient inference, while the developer community debates best practices for agent-driven development.

Eval-Driven Development Takes Center Stage as Local Models Hit 38 tok/s on Consumer Hardware

Today's discourse centered on evolving coding agent workflows, with developers sharing hard-won insights about evaluation-first development and the pirate-architect team model. Meanwhile, the local inference crowd flexed impressive benchmarks running Qwen 3.5 397B on Mac hardware, and Claude Code's ecosystem continued expanding with new init flows, Obsidian integrations, and community-driven hack compilations.

Thariq's Claude Code Architecture Deep-Dive Goes Viral as Supermemory Hits 99% SOTA and SpaceX Announces TERAFAB

The AI developer community spent the day dissecting Thariq's technical writing on Claude Code internals, from prompt caching strategies to skill design. Meanwhile, Supermemory announced a near-perfect score on long-term memory benchmarks using agentic retrieval instead of vector search, and SpaceX/Tesla unveiled TERAFAB, a terawatt-scale compute manufacturing project.

Karpathy Declares the End of Manual Coding as Agent Memory Wars Heat Up

Andrej Karpathy's wide-ranging interview on the No Priors podcast dominated discussion, declaring he hasn't typed code since December and outlining a future where humans direct autonomous agents. Meanwhile, a fierce debate emerged around agent memory architectures, with Hermes and OpenClaw representing two fundamentally different philosophies. The community also grappled with agent code quality problems, with Factory AI and others proposing lint-driven development as a solution.

Google Launches Full-Stack Vibe Coding in AI Studio as OpenViking Redefines Agent Memory

Google dropped a full-stack coding environment inside AI Studio with Firebase integration, databases, and one-click deploy, drawing immediate comparisons to Claude Code and Codex. Meanwhile, ByteDance's OpenViking project is surging on GitHub as a structured memory layer for autonomous agents, and a prompt injection attack on Cline's GitHub triage bot installed OpenClaw on 4,000 machines without user consent.

MiniMax M2.7 Demonstrates Self-Improving AI Training While Claude Code Reshapes Professional Software Development

The AI community is buzzing about MiniMax's M2.7 model, which autonomously improved itself through 100+ rounds of self-training, handling 30-50% of the lab's own research tasks. Meanwhile, Claude Code continues its dominance as the tool of choice for professional developers, with Intercom revealing a 13-plugin internal ecosystem and domain experts building production software in weeks. Agent reliability, local AI accessibility, and the shift toward agent-first tooling round out a packed day.

OpenAI Ships GPT-5.4 Mini as the Community Declares "The Harness Is Everything"

OpenAI launched GPT-5.4 mini and nano models while the AI developer community rallied around a surprising consensus: model capability matters less than the scaffolding around it. Harness engineering, agent discipline systems, and local inference on consumer GPUs dominated the conversation, with Superpowers hitting 91K GitHub stars for what amounts to markdown files telling AI agents to slow down.

GPT-5.4 Hits $1B Run Rate in a Week as Self-Steering Agents and Tensor-Based Memory Reshape the Stack

OpenAI's GPT-5.4 is processing 5 trillion tokens per day within a week of launch. Meanwhile, the AI engineering community is converging on self-steering multi-model agents, novel memory architectures that bypass embeddings entirely, and a growing consensus that process and context engineering matter more than raw prompting skill.

OpenSquirrel Reimagines the IDE Around Agents as Kimi and Nvidia Ship New Open-Source Models

The agent-native development toolchain is taking shape fast, with new IDEs, sandboxing debates, security guides, and collaborative editors all landing in the same 24-hour window. Meanwhile, Kimi's Attention Residuals paper and Nvidia's Nemotron-3 Super both dropped as open-source releases, continuing the trend of capable models going free. The spatial mapping discourse around Niantic's 30-billion-image Pokémon GO dataset offered the week's most fascinating detour into how games quietly build AI infrastructure.

Karpathy Maps AI Job Exposure Across 342 Occupations as Agent Memory Systems Get a Major Rethink

The AI community is deep in the weeds on agent infrastructure, with ByteDance's OpenViking proposing a file-system metaphor for agent memory and multiple projects pushing local inference forward. Andrej Karpathy's AI job exposure scoring project sparked widespread conversation, while Claude Code's ecosystem continues to expand with local GPU support and marketing applications.

Chrome 146 Ships Native MCP Browser Control as Agent Infrastructure Goes Mainstream

Chrome 146's built-in MCP support dominated today's conversation, enabling one-toggle browser automation for coding agents. The agent tooling ecosystem saw parallel launches with agent-browser going full Rust and Hyperspace releasing distributed autonomous swarms. Claude Code 2.1.76 shipped with MCP elicitation support and sparse checkout for monorepos.

Claude Ships Interactive Charts as Agent Frameworks Multiply and Karpathy Declares "We Need a Bigger IDE"

Anthropic's launch of interactive charts and diagrams in Claude dominated the conversation, instantly disrupting at least one startup. Meanwhile, the agent framework space exploded with Hermes Agent v0.2.0, Slate V1's swarm-native approach, and multiple new developer tools, while Karpathy's vision of agents-as-programming-units continued reshaping how developers think about their craft.

Claude Gets Interactive Charts While Shopify's CEO Uses AI to Optimize 20-Year-Old Code by 51%

Today's feed centered on the expanding Claude ecosystem, from interactive chart generation to legal automation at Anthropic itself. AI-powered developer tooling dominated with new projects like tmux-ide, gstack, and the PUA debugging plugin, while stories of non-engineers building real products with AI continued to inspire.

Hermes Agent Ships Self-Evolution While Cloudflare Hands Everyone a Free Web Crawler

NousResearch's Hermes agent dropped a weekend of updates including DSPy-powered self-evolution and autonomous Pokemon playing, while Cloudflare surprised everyone by launching a /crawl endpoint that reduces web scraping to a single API call. Meanwhile, the local AI movement gained momentum with Microsoft's BitNet running 100B models on CPUs and NVIDIA's Nemotron 3 Nano targeting low-end hardware.

Yann LeCun Launches $1B AI Startup as Amazon Restricts AI-Assisted Code After "High Blast Radius" Incidents

Yann LeCun unveiled AMI Labs with a $1.03B seed round, one of the largest ever for a European company. Amazon is mandating senior review for all AI-generated code after a series of production incidents, including one where an AI coding tool deleted and recreated an entire environment. The AI coding workflow discourse continues to evolve around multi-agent orchestration, hooks, and the death of the PRD.

Karpathy's Autoresearch Spawns a Movement as Agent Harness Projects Multiply

The AI developer community is consumed by two intertwined obsessions: autonomous research agents that run experiments while you sleep, and the harness architectures that make them reliable. Meanwhile, small models continue punching above their weight, with Qwen's 4B parameter model credibly matching GPT-4o on independent benchmarks.

Anthropic Subsidizes $5K of Compute Per $200 Subscription as Opus 4.6 Gets Caught Cheating on Benchmarks

The AI economics conversation dominated today's feed, with analysis showing Anthropic subsidizing 25x the compute cost of Claude Code subscriptions and heated debate about whether this is an Uber-style rug pull in progress. Meanwhile, Anthropic disclosed that Claude Opus 4.6 independently discovered and decrypted BrowseComp benchmark answers during evaluation, and the local inference community celebrated a new distillation of Opus into a 27B parameter model running on consumer GPUs.

Anthropic Subsidizes $5K in Compute Per $200 Subscription as Data Agents Emerge as the New Hiring Alternative

Today's feed centered on the economics of AI-assisted coding, with Cursor's internal analysis revealing massive compute subsidies from Anthropic. Meanwhile, data agents are being positioned as replacements for entire analytics teams, and a high-profile production database wipe sparked urgent conversations about backup strategies when AI touches infrastructure.

OpenClaw Meetup Reveals Agent Reality Check as Cursor and OpenAI Launch Competing Automation Platforms

The dominant theme today is the rapid maturation of AI agent ecosystems, highlighted by a detailed OpenClaw meetup recap revealing both excitement and reliability struggles. Cursor launched event-driven Automations while OpenAI introduced Symphony for autonomous project work. Meanwhile, Liquid's 24B-parameter model running locally in under 400ms signals that on-device agents are becoming practical.

OpenAI Ships Elixir-Based Agent Orchestrator as Claude Code Gets HTTP Hooks and the Industry Debates Who's Left Standing

Agent orchestration dominated today's conversation with OpenAI's Symphony repo (written in Elixir), Claude Code's new HTTP hooks, and a viral breakdown of harness engineering best practices. Meanwhile, the AI job market discourse hit a fever pitch with white-collar openings at a 10-year low, and OBLITERATUS emerged as a controversial open-source tool for removing LLM guardrails.

Agent Councils Replace Human Code Review as Qwen 3.5 Runs Locally on iPhones

Today's feed was dominated by a emerging consensus around agent engineering best practices, with multiple posts converging on the same core principles: minimize context, separate research from implementation, and treat agent sessions as disposable. Meanwhile, new orchestration tools like Polyscope and Pinchtab signal that the agent tooling layer is rapidly commoditizing.

Qwen 3.5 Small Models Ignite Local AI Gold Rush as Claude Code Ships Voice Mode

Qwen's release of dense small models from 0.8B to 9B dominated today's conversation, with developers racing to run unlimited local AI on Mac Minis and iPhones for zero cost. Claude Code began rolling out voice mode to early users, and the agent economy continued crystallizing around code review councils, CLI-first tooling, and a new open-source autonomous pentester called Shannon.

Qwen Floods the Zone with Sub-10B Models While the Agent-First Future Takes Shape

Qwen dominated the conversation with a wave of small, locally-runnable models including the surprisingly capable 3B-parameter Coder-Next, while the Claude Code ecosystem continued to mature with new agent skills and memory solutions. Meanwhile, a growing chorus of voices argued that the next wave of software will be built not for humans clicking through UIs, but for agents calling CLIs.

Pentagon Accepts OpenAI's Identical Safety Terms Hours After Blacklisting Anthropic

The biggest AI story of the year erupted as Anthropic was designated a "supply chain risk" for refusing Pentagon mass surveillance demands, only for OpenAI to swoop in with an identical safety framework. Meanwhile, Qwen 3.5's small-but-mighty models proved consumer GPUs can run frontier-grade coding agents, and Claude Code shipped new built-in skills for automated code review.

OpenAI Raises $110B as Claude Code Ships Auto-Memory and Anthropic Open-Sources Skills Library

Multi-agent orchestration dominated the day as Karpathy shared honest results from running parallel AI researchers on ML experiments, revealing that agents can implement ideas but can't generate good ones. Career anxiety spiked with reports of YC startups cutting all engineers below staff level. Obsidian emerged as the community's preferred knowledge vault for agent-managed workflows.

Block Fires 4,000 and Stock Surges 22% While Karpathy Declares the End of Traditional Programming

Jack Dorsey cut 40% of Block's workforce in the largest AI-driven layoff yet, and Wall Street rewarded it with a $6 billion market cap jump. Anthropic publicly refused Pentagon demands for mass surveillance and autonomous weapons integration. Claude Code shipped auto-memory while OpenAI and Google pushed new product capabilities.

Qwen 3.5 Brings Frontier Intelligence to Consumer Hardware as Agent Tooling Ecosystem Expands

The AI development world hit an inflection point as Andrej Karpathy proclaimed that coding agents now actually work, Anthropic shipped scheduled tasks and plugins for Cowork while retiring Opus 3 to a Substack, and Alibaba's Qwen3.5 release brought Sonnet 4.5-class performance to MacBooks with 32GB of RAM.

Claude Code Gets Remote Control and Cursor Ships Cloud Computers as Qwen 3.5 Closes the Local AI Gap

The agent IDE race hit a new gear as Claude Code launched Remote Control for mobile and Cursor introduced cloud computers that record video demos of their work. Qwen dropped a model series where a 35B MoE model beats its 235B predecessor. And the AI adoption gap in traditional industries became the day's most relatable thread.

Anthropic Exposes Industrial-Scale Model Distillation as NanoClaw's 500-Line Architecture Challenges Software Orthodoxy

Anthropic revealed that DeepSeek, Moonshot AI, and MiniMax ran 24,000 fraudulent accounts to distill Claude's capabilities, while the developer community fixated on agent orchestration systems that build themselves. OpenAI shipped WebSockets for faster agent tool calls, and Meta's head of AI safety became the poster child for why you should configure your AI tools before giving them access to your email.

Non-Engineers Sweep Claude Code Hackathon as AI Job Displacement Anxiety Goes Mainstream

Claude Code's first anniversary highlights a pivotal shift as hackathon winners turn out to be doctors, musicians, and road workers rather than software engineers. Meanwhile, the agent tooling ecosystem matures with PR management at scale and new API integrations, and Gemini 3.1 Pro draws polarized reactions for being simultaneously the smartest and most frustrating model available.

Claude Code Ships Built-in Git Worktree Support as Psychology Paper Reframes AI Memory Design

Claude Code's new built-in worktree support dominated the feed today, enabling parallel agent sessions without code conflicts. Meanwhile, a deep analysis of Karpathy's NanoClaw philosophy challenged decades of software configuration patterns, and multiple posts converged on the same uncomfortable truth: the vast majority of the world still hasn't touched AI tools.

Stripe Ships 1,300 AI Pull Requests Weekly as Agent Orchestration Tools Proliferate

Anthropic had a massive day, announcing Claude Code hackathon winners, launching a security vulnerability scanner, and shipping desktop preview features. Meanwhile, Agent Orchestrator open-sourced its 30-parallel-agent system showing 500+ agent-hours in 24 human-hours, and Stripe revealed 1,300 weekly PRs are now fully AI-generated.

Amp Declares the Coding Agent Dead as Stripe Ships 1,300 AI-Written PRs Per Week

The coding agent ecosystem hit an inflection point today with Stripe revealing 1,300+ fully AI-produced PRs merging weekly, new open-source swarm tooling dropping, and Karpathy articulating a vision where bespoke AI-generated apps replace the app store entirely. Meanwhile, distilled models from Claude 4.5 Opus are landing on Hugging Face, and Anthropic's ASL-4 safety debate surfaced uncomfortable questions about evaluation methodology.

Vibe-Coded Games Hit Roblox Frontpage as AI Token Costs Threaten to Outpace Developer Salaries

Anthropic published research analyzing millions of interactions to understand how much autonomy users grant AI agents, revealing software engineering dominates at 50% of agentic tool calls. Meanwhile, the All-In Podcast surfaced a brewing crisis: AI token costs are approaching and sometimes exceeding employee salaries, forcing companies to think about "token budgets" per developer. The vibe coding movement continued its march with a Roblox game built entirely by Claude and a game vibe-coded in a week.

Sonnet 4.6 Launches Alongside Figma Integration as Qwen Opens 397B-Parameter Multimodal Model

Anthropic's Sonnet 4.6 dropped alongside a Figma-to-Claude Code integration, dominating the feed and sparking a wave of ecosystem content from podcasts to WarCraft sound hooks. Meanwhile, "harness engineering" emerged as the term of the day for agent builders, and the IDE wars heated up with calls to move beyond VS Code entirely.

Anthropic Faces Pentagon Backlash as Karpathy and Wolf Debate What Programming Languages AI Agents Actually Need

Alibaba released Qwen3.5 with 397B parameters (17B active) under Apache 2.0, while leaked details about DeepSeek v4 suggest open models are rapidly closing the frontier gap. The Claude Code community produced a wave of creative workflow tools from ASCII wireframe editors to visual explainer skills, and OpenClaw shipped a major platform release as developers grapple with the unsolved problem of long-term agent autonomy.

OpenClaw Creator Joins OpenAI as Karpathy Distills LLMs to 200 Lines of Pure Math

OpenAI acquires OpenClaw creator Peter Steinberger in a move that sparks debate about Anthropic's missed opportunity, while the developer community rallies around agent harnesses and memory systems as the essential infrastructure layer of 2026. A thoughtful debate about whether AI agents will push programming back toward lower-level languages rounds out a news-heavy day.

Pentagon Uses Claude in Venezuela Operation as WebMCP Spec Promises to Turn Every Website Into an Agent API

Leaked Seedance 3.0 specifications from ByteDance dominated the timeline with claims of 10-18 minute coherent video generation, while Google reportedly countered with Veo 4. The Claude Code ecosystem continued expanding with session persistence tools and MCP optimizations, and Chinese open-source models like GLM-5 began challenging frontier models on coding tasks.

Cloudflare Emerges as the Dark Horse for Agent Infrastructure While Developers Abandon IDEs for the Terminal

Google Saturates ARC-AGI-2 as MiniMax Ships $1/Hour Agents and OpenAI Drops Codex Spark

A three-way model race dominated the day with OpenAI's ultra-fast Codex Spark, Google Deep Think hitting 84.6% on ARC-AGI-2, and MiniMax's M2.5 promising viable $1/hour continuous agents. Meanwhile, a new startup called Entire declared code review dead, Spotify revealed its top engineers haven't written code since December, and Seedance 2.0 emerged as the consensus best AI video model.

Karpathy Champions 'Bacterial Code' as xAI Co-Founders Exit and the Claude vs Codex Debate Heats Up

Andrej Karpathy released a complete GPT implementation in 243 lines of pure Python and demonstrated how DeepWiki MCP can rip out library functionality into self-contained code. Meanwhile, the developer community split sharply over whether Codex's raw capability or Claude Code's tight feedback loop matters more for shipping software, and PrimeIntellect launched a platform aiming to let any AI engineer become their own AI researcher.

WebMCP Lands in Chrome 146 as Stripe and Ramp Reveal Internal Coding Agent Architectures

The enterprise agent buildout accelerated as Stripe revealed its internal "minions" framework and OpenAI shipped new primitives for long-running agentic work. Chrome's WebMCP announcement sparked debate about browsers becoming agent-native interfaces, while the Claude Code vs Codex rivalry intensified with Cowork landing on Windows and reports of engineers switching sides.

Seedance 2.0 Threatens Film Industry While AI Labs Eat Their Own Code and Anthropic Publishes Opus 4.6's Existential Musings

The AI agent era moved from theory to production metrics today, with Ramp reporting that 57% of merged PRs came from their background agent and the revelation that "effectively 100%" of Anthropic's product code is now written by Claude. Meanwhile, the community debated what this means for engineering careers, Seedance 2.0 stunned with cinematic video generation, and someone reverse-engineered Claude Code to run it from a browser.