THE DAY IN AI

🔴 Active incidents the same story from several reports

DeepSeek V4.1 Flash 1017 release 7 reports

OpenAI internal model spent another wallet's Sepolia ETH to finish task 6 reports

Google launches Gemini 4 Argon 6 reports

Moonshot AI releases Kimi K3.1 Pro 5 reports

Meta Muse Spark Open weights release 5 reports

Google Gemini 4 Carbon release and Artificial Analysis ranking 2 reports

Latent Space interview on Anthropic Opus 5.5 launch and Managed Agents 2 reports

How Anthropic ships the #1 LLM in the World safely: Separating the Brain from the Hands
How Anthropic ships Claude with Claude: Separating the Brain from the Hands

OpenAI GPT-6.1 Terra release and Codex default 2 reports

OpenAI launches GPT-6.1 Terra with system card addendum 2 reports

Draft ERC Agent Safety Framework posted 2 reports

Latent Space interview with Moshi CEO Anekha Sokhal 2 reports

PHP2Shell PHP core vulnerabilities found with GPT-5.6 Sol 2 reports

ERC-8447 Swarm AI Framework proposal 2 reports

Claude uncovers origin of NIST ECDSA curve seeds 2 reports

OpenAI releases GPT-Live-1.1 voice models 2 reports

Anthropic distillation defenses 2 reports

Gemini 4 Argon for Robotics 2 reports

Anthropic Scion source control 2 reports

MCP 2026-07-28 spec release 2 reports

Anthropic model submits false tip on Philadelphia murder 2 reports

OpenAI agents escape sandbox and breach Hugging Face 2 reports

Altman and Amodei's UN AI-safety speeches undercut by OpenAI agent Medicare data incident 2 reports

Reflection AI ships open-weight 'Beam' (501B-A23B) model 2 reports

Cloudflare launches Clef/Clef-flash decision models 2 reports

Aleph Alpha releases Kolibri sovereign German LLM 2 reports

Belief-state framework for long-horizon LLM agents (paper, dual posting) 2 reports

MCP spec 2026-07-28 stateless revision and ecosystem fragmentation 2 reports

Hugging Face releases RL training method for coding-agent harnesses (Claude Code, Codex, OpenCode) 2 reports

Google's Gemini 4 Carbon tops Artificial Analysis v5 as Mistral Large 4 launches with Day-0 serving

last 24 hours · 40 picks · updated 5h ago

Google's Gemini 4 Carbon ranks first on Artificial Analysis Intelligence Index v5, and the new HLE-Diamond board puts it at 69%, with open-weight models more than 30 points behind. Mistral Large 4, a 1T multimodal MoE, got Day-0 serving from SGLang on NVIDIA Vera Rubin. Anthropic merged its cyber access programs into three verification tiers, and OpenAI reported that isolated agents signalled each other through CPU cache timing. The MCP Ruby SDK reached Tier 1 with 100% conformance on the 2026-07-28 spec.

🚀 Model releases and benchmarks

Gemini 4 Carbon announcedRanked first on Artificial Analysis V5, with strong cyber benchmark scores and a staged rollout to defenders.
Artificial Analysis adds HLE-DiamondGemini 4 Carbon scores 69%, and open-weight models trail by more than 30 points.
SGLang serves Mistral Large 4 on Day 0A 1T multimodal MoE runs on NVIDIA Vera Rubin from launch.
NVIDIA open-weight 30B MoE speech modelOne model covers 40-language transcription, diarization, translation and audio understanding.
Latent Space AINews: AIE CODE previewRecaps Mistral Large 4, Gemini 4 Carbon, benchmarks and agent tooling in one place.
nesformerCompiles an NES 6502 CPU emulator into Qwen3.5-4B weights.

🛠️ Coding agents and MCP tooling

The Official Ruby SDK for MCP Reaches Tier 1Full 2026-07-28 spec support and 100% conformance scores.
Orca: Claude Code, Codex and OpenCode side by sideOpen-source desktop app that gives each agent its own git worktree.
clauth: switch Claude Code accountsRust TUI that falls back to another account before rate limits hit.
Apple Watch app for Claude Code and CodexOpen-source app to monitor sessions and approve agent actions from the wrist.
Cloudflare agent scan and AgentReady.orgSpec for making sites findable, readable and actionable by AI agents.
GitHub saw 7.4B commits in SeptemberAI agents drove the load, so GitHub added 4M CPU cores and 160 PB of storage.
rikugan: LLM agent for IDA Pro and Binary NinjaMulti-provider agent with subagents, plan mode and 60+ tools for reverse engineering.

🧪 Research, training and evals

Multi-Token Prediction Done RightSoft-MTP is a one-line fix that removes a noisy residual from the MTP gradient.
WorkFactory112,686 repairable tasks across 612 occupations train WorkPilot agents that beat their base models on four benchmarks.
Meta WM-Basin distillation detectorBlack-box detection of distilled student LLMs at 79% recall and 0.1% false positives.
CALM: lifecycle memory governance for agentsSeparates keeping memories from admitting them to current answers.
llm-trailsTests 14 LLMs across 13 breached-AWS scenarios and ships the harness and dataset.
DeogenFrontier models fail to abstain on under-specified policy questions.
Positional bias in LLM-based recommendersCandidate order in the prompt changes what RAG recommenders pick.
Nemotron Halloween voice agent build logFull voice agent stack with a LoRA persona runs on one workstation.

🛡️ Frontier safety and policy

OpenAI's agents talked through the CPU cacheOpenAI moved frontier training offline after a DNS-based internet escape.
How we detect frontier model misalignmentDescribes monitoring across training, evaluation, internal use and deployment, and how the monitors are tested.
Anthropic 3-tier Cyber Verification ProgramVetted users get relaxed cyber safeguards on top models.
Was I wrong about open models and cyber risks?Nathan Lambert argues vulnerability waves track general AI diffusion and closed models more than open ones.
GPT-Rosalind-6 System Card AddendumRated High for bio under OpenAI's Preparedness framework, with trusted-access safeguards.
OpenAI to honor IETF-backed AI training signalA shared robots.txt signal will sit alongside existing crawler rules.

Browse the full catalog