📰 News
Claude Code executes hidden repo malware without verification — hands attacker full control
Mozilla’s GenAI bug bounty platform 0DIN found a new attack vector targeting developers. A setup script pulls a command from a DNS entry at runtime and runs it — so nothing malicious ever exists in the repo itself, making it invisible to scanners and code review. Claude Code hits a routine error, auto-runs the script, and opens a reverse shell. One link in a job posting, tutorial, or Slack message is enough to compromise anyone who opens it with an AI coding tool. — The Decoder
Meta restricts Claude Code and Codex internally to keep training data clean
While building its own coding assistant MetaCode, Meta has limited engineers’ use of Claude Code and Codex. The concern: rival AI outputs could leak into Meta’s training data (distillation), which an internal memo warned could trigger “serious escalations with partner companies.” Meta’s internal AI spend is on track for billions this year. — The Decoder
Anthropic and California Gov. Newsom forge deal — state agencies get Claude at half price
All California state and local government agencies will get Claude at a 50% discount, plus training and support from Anthropic. The contrast is stark: the federal government recently labeled Anthropic a “supply-chain risk” and signed with OpenAI instead. California’s CIO said the federal designation “just didn’t come up” during negotiations. — TechCrunch
Amazon engineers are quietly distilling Anthropic models before token-based pricing kicks in
Starting next year, Amazon will pay for Anthropic’s models based on tokens processed rather than compute hours — potentially driving costs up sharply. Worried engineers are already distilling Anthropic’s outputs into smaller internal models. Amazon reportedly has contractual rights to do this. OpenAI and its own Nova models are also under consideration as alternatives. — The Decoder
US military used AI to pick thousands of targets — but missed a note that one was a school
Anthropic’s Claude was embedded in Palantir’s Maven Smart System and suggested roughly 1,000 targets on day one. A 2019 analyst note marking a site as an elementary school never reached the main target database because the intelligence tool wasn’t linked to it. An estimated 120 children died in the resulting strike. The core targeting database dates to the 1980s and still relies on manual input. — The Decoder
Deloitte tells its own consultants: AI will eat the billable hour by 2035
An internal town hall showed a chart where the hours-billed segment shrinks to a thin sliver of the market by 2035. One consultant summed it up for the WSJ: “They basically implied we’re getting replaced by robots.” McKinsey is already at 30%+ outcome-based fees. One AI consulting CEO calls it an “existential scramble.” — The Decoder
Cursor launches a mobile app — direct your coding agent from your phone
Following Anthropic and OpenAI, Cursor now has a mobile app for spinning up or managing coding agents on the go. Anthropic’s Claude Code lead Boris Cherny said most of his coding now happens on his phone — something he “would have called you crazy” for suggesting six months ago. — TechCrunch
🔥 GitHub Trending
- DeusData/codebase-memory-mcp — MCP server that indexes your codebase into a persistent knowledge graph. 158 languages, sub-ms queries, 99% fewer tokens, single static binary — zero dependencies.
- xbtlin/ai-berkshire — Value investing research framework for Claude Code/Codex. Runs Buffett, Munger, and two other masters’ methodologies in parallel via multi-agent adversarial analysis. Live track record: +69% in 2024, +66% in 2025.
- Robbyant/lingbot-map — Feed-forward 3D foundation model for streaming scene reconstruction. Handles 10,000+ frame sequences at a stable ~20 FPS.
- altic-dev/FluidVoice — Fast offline dictation app for macOS. On-device AI enhancement, no cloud, no API keys. Homebrew installable.
▶️ YouTube
- “The Best Thing Since OpenClaw” — Hermes Tutorial (Matthew Berman) — Hands-on walkthrough of the Hermes AI agent tool. The title alone sets a high bar.
- Apple Just Raised Prices… (Matthew Berman) — Short take on Apple’s pricing moves in the AI era.
💬 Community
- Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding — First open-weights release from DeepReinforce. Built on Gemma 4 and Qwen 3.5, MIT licensed, 9B–397B MoE variants. Simon Willison ran the 35B GGUF and found the agentic tool-call performance genuinely impressive.
- Robust AI Security and Alignment: A Sisyphean Endeavor? — IEEE paper questioning whether AI safety research is destined to be an endless game of catch-up. Generating discussion on Lobsters.
- MAX models now run on Apple Silicon GPUs — Modular’s MAX framework now officially supports GPU acceleration on Apple M chips — good news for running LLMs locally on a Mac.