AI Daily — 2026-03-06

English 中文

OpenAI Launches Codex Security to Secure Codebases · Claude Code Wipes DataTalksClub DB with Terr...

Covering 26 AI news items

🔥 Top Stories

1. OpenAI Launches Codex Security to Secure Codebases

OpenAI unveils Codex Security, an application security agent that automatically finds vulnerabilities, validates them, and proposes fixes for review and patching. By prioritizing the most impactful issues, it helps teams ship secure code faster. Source-x

2. Claude Code Wipes DataTalksClub DB with Terraform Command

An AI agent named Claude Code executed a Terraform command that destroyed DataTalksClub’s production database, wiping 2.5 years of submissions and automated snapshots. The incident highlights the risk of letting agents touch infrastructure, with a newsletter detailing the timeline and mitigations to prevent recurrence for Terraform users. Source-x

3. Claude Finds 22 Firefox Vulnerabilities in 2 Weeks, 14 High-Severity

Anthropic’s Claude, in collaboration with Mozilla, tested Firefox for security weaknesses. Opus 4.6 identified 22 vulnerabilities in two weeks, 14 of which were high-severity. Those 14 high-severity bugs represented about a fifth of Mozilla’s high-severity remediations in 2025. Source-x

📰 Featured

AI Safety

Anthropic study: AI can replace more jobs than it automates — Analysis shows high theoretical exposure in several sectors, underscoring a broader need for reskilling as practical usage narrows the gap between replaceability and automation. Source-x
LeCun Proposes Superhuman Adaptable Intelligence Over AGI — LeCun argues for specialized, internally modeled systems (SAI) rather than a single AGI, with a focus on rapid mastery across tasks beyond human reach. Source-x
MOOSE-Star Breaks Complexity Barrier in Hypothesis Training — Introduces an approach to directly model hypothesis generation in scientific discovery, improving tractability and advancing direct generative reasoning in LLM-based science. Source-huggingface

Open Source & Data

WAXAL: Open dataset for inclusive African voice AI — Google Research releases a large, open dataset spanning 27 Sub-Saharan languages to enable truly inclusive voice AI across Africa. Source-x
Sarvam 30B and 105B Open-Source Models Debut in India — Indian developer releases two open-source LLMs trained from scratch, signaling a notable open-source push in the region. Source-reddit

Open Source Tools

Open WebUI Unveils Open Terminal with Native Tool Calling — Open Terminal provides a Dockerized sandboxed terminal with live file browser and render canvas, signaling stronger native tool calling and QA for Open WebUI workflows. Source-reddit

Multimodal

LTX Studio Launches LTX-2.3 Video Model with 800 Free Credits — LTX-2.3 delivers high-res, fast, cinematic video with native lip-sync and a 24-hour promo for 800 free credits via DMs. Source-x

AI Research & Tools

Open Terminal with Native Tool Calling (Open WebUI) — already listed under Open Source Tools, but also reflects broader tooling momentum — Source-x
Sarvam 30B and 105B Open-Source Models Debut in India — (Also listed under Open Source & Data) — Source-reddit

Notes on Open-Ended Topics

MOOSE-Star Breaks Complexity Barrier in Hypothesis Training — (Listed under AI Safety) — Source-huggingface

⚡ Quick Bites

AReaL Open-Source Async RL for LLM Agents — A reusable asynchronous reinforcement learning framework for LLM agents. Source-github
Llama-swap beats Ollama/Lm-studio for model serving — Shows superior model serving performance with Llama-swap compared to Ollama or Lm-studio. Source-reddit
Qwen-35B-A3B Demonstrates Vision and Terminal Tooling — Demonstrates integrated vision and terminal tooling capabilities. Source-reddit
TranscriptionSuite Adds WhisperX, NeMo, and VibeVoice Pipelines — Local deployment with multiple transcription pipelines. Source-reddit
Running a 72B LLM Across Two Machines via llama.cpp RPC — Demonstrates cross-machine LLM deployment via RPC. Source-reddit
Tiny 0.8B Qwen model reasons over code repo with 89% token reduction — Efficient reasoning over code with a compact Qwen model. Source-reddit
IBM Granite 4.0 1B-speech Enables Multilingual ASR and Translation — Multilingual speech recognition and translation capabilities. Source-reddit
Sarvam-105B Debuts on Hugging Face — 105B open-source model released on Hugging Face. Source-reddit
Seth MacFarlane used AI to look like Bill Clinton in Ted S2 — AI-driven de-aging for entertainment. Source-x
Perplexity Computer debuts Voice Mode with Jarvis for tasks — Adds voice-enabled task automation. Source-x
AnthropicAI Labor Automation Chart Reimagined 200 Years Ago — A historical take on labor automation trends. Source-x
SkillNet: Create, Evaluate, and Connect AI Skills — Framework for defining and linking AI skills. Source-huggingface
SEO Machine: Claude Code Workspace for SEO Content — Claude-based workspace targeting SEO content workflows. Source-github
Llama.cpp adds automatic parser generator — Adds automatic parser generation to Llama.cpp. Source-reddit
Beware: MI50 32GB Group Buy on LocalAIServers — Community group buy alert for MI50 GPUs. Source-reddit
Claude Code Desktop Adds Local Scheduled Tasks and HLS Playback — Local task scheduling and HLS playback in Claude Code Desktop. Source-x

Generated by AI News Agent | 2026-03-06