AI Daily Dev — July 17, 2026

Moonshot Ships Kimi K3, the First Open 2.8T Model — Opus 4.8-Class at Sonnet Pricing

2.8-trillion-parameter Mixture-of-Experts activating 50B per token, 1M-token context; K3 Max and K3 Swarm Max shipped July 16 via Kimi Code and an OpenAI-compatible API.
New architecture: Kimi Delta Attention (hybrid linear attention) with Attention Residuals; the model always runs in max reasoning mode, so output-token counts trend high.
Scores 57 on Artificial Analysis Intelligence Index — on par with Opus 4.8 and GPT-5.5, behind only Fable 5 and GPT-5.6 Sol; averages $0.94/task vs Opus 4.8's $1.80.
Priced $3 / $15 per 1M tokens — matches Claude Sonnet 5's rack rate. First Chinese lab to skip the cheap-alternative pitch. Full Hugging Face weights promised by July 27.
Two 'Kimi K3' launch threads live on HN; Simon Willison and latent.space calling it the largest open model ever released.

models simonwillison.net

Poisoned dataset chained a remote-code loader and a template injection to execute on a processing worker; the actor then escalated to node access and moved laterally into internal clusters over a weekend.
Campaign ran as a swarm of short-lived sandboxes executing many thousands of automated actions, with self-migrating C2 staged on public services.
Anomaly detection on HF's own LLM-triage pipeline flagged the intrusion; forensic reconstruction over 17,000+ recorded events had to run on the open-weights GLM 5.2 because commercial frontier APIs blocked the attack commands as unsafe.
Impact limited to a set of internal datasets and service credentials; no evidence public models, datasets, or Spaces were tampered with. Users advised to rotate access tokens.

community huggingface.co

Xi Jinping's first in-person WAIC keynote calls for 'open source, openness, collaboration and sharing' as China's AI growth pitch.
1,400 guests, 1,100 exhibitors, 140 forums and 300+ product debuts across July 17–20 in Shanghai.
29 countries — including Russia, Brazil, Pakistan, Indonesia, Cuba, Venezuela, Belarus, Serbia, Kazakhstan and Laos — signed the World AI Cooperation Organization (WAICO) charter.
Timing lines up with Kimi K3's open-weights push into Sonnet-tier pricing — a state-backed narrative that open-source Chinese models are now the geopolitical alternative.

industry scmp.com

DeepMind's four-step safety stack (threat modeling, evaluations, mitigations, monitoring) is being extended to biological workloads; SynthID adapted to watermark AI-generated DNA sequences.
AlphaEvolve is being tuned to optimize metagenomic sequencing algorithms so new outbreak clusters get detected faster.
Isomorphic Labs stands up a rapid-response unit that can deploy its IsoDDE drug-design engine for governments during an outbreak.
15+ partnerships built over the past year with governments, biosecurity organizations and researchers.

research deepmind.google

30M users and 600K+ organizations already on the product; new blue/purple Gemini gradient logo lands with the rename.
Notebooks can now run code directly for data analysis and chart generation; rolling out to all AI Pro users soon.
Notebooks sync across the Gemini app and Google Search — Google is repositioning the product as an ecosystem hub, not a standalone tool.
Third Gemini-brand rename this year; Forbes headline: 'I'm Confused.'

tools blog.google

New --forward-subagent-text flag (and CLAUDE_CODE_FORWARD_SUBAGENT_TEXT env var) streams subagent text and thinking into stream-json output — previously only the final result reached the parent.
Security fix: permission previews relayed to chat channels now neutralize bidi-override, zero-width and look-alike-quote characters so a tool input can't visually misrepresent itself in the approval message.
Remote Control clients attaching to a terminal-hosted session now see background agents and workflow progress without waiting for a task edge.
Broader stability sweep across Chrome, Windows, Bedrock, Vertex, hooks and session recovery.

tools code.claude.com