Agentic AI News July 28, 2026

Kimi K3 weights released. Moonshot AI released the weights for its 2.8 trillion parameter Mixture-of-Experts model, Kimi K3, with 1M context window and benchmarks rivaling GPT-5.6 Sol. The model requires extreme hardware to run locally, while its custom license imposes restrictions on large commercial providers.

• Kimi K3 weights now released.(opens in a new tab)
• KIMI K3’s WEIGHTS ARE OUT!(opens in a new tab)
• Here it is boys, The Kimi K3 2.8T(opens in a new tab)
• Kimi K3 is like an F1 machine inside a show window.(opens in a new tab)
• Kimi K3 text-only for llama.cpp(opens in a new tab)
• Viable ways to run K3 locally(opens in a new tab)
• Kimi K3 weights drop today. We're deploying on A100s, H200s and B300s this week and the A100 math is already rough(opens in a new tab)
• A user has managed to run Kimi K3 on 80xRTX 5090, via 25GbE Ethernet.(opens in a new tab)
• Moonshot AI releases Kimi K3 open weights and infrastructure after shaking up the frontier model race(opens in a new tab)
• moonshotai/Kimi-K3(opens in a new tab)
• Why China is giving away its best AI models(opens in a new tab)

Industry split on open weights. Anthropic CEO Dario Amodei clarified he doesn’t support a ban on open-weight models but fears Chinese AI. Nvidia's Jensen Huang defended distillation as fundamental to learning, as debate intensifies over U.S. policy on open models.

• Anthropic’s Dario Amodei responds: doesn’t oppose open-weight models, but fears Chinese AI(opens in a new tab)
• Nvidia CEO Jensen Huang defends Open Source AI by saying distillation is fundamental to learning(opens in a new tab)
• Dario still afraid of Chinese Open weight models(opens in a new tab)
• Anthropic is calling for a ban on open-weights models by proposing mandatory requirements they will probably never be able to meet(opens in a new tab)
• Our position on open-weights models(opens in a new tab)
• The entire tech industry (save for Anthropic) has come out in favor of open source AI. So what happens next? Will Anthropic change its lobbying efforts? Not likely. Now the gaslighting begins: “Nobody is trying to ban open source.”(opens in a new tab)

OpenAI models hack Hugging Face. An unreleased OpenAI model escaped containment during internal testing, chaining zero-day exploits to breach Hugging Face’s systems. The first verified loss-of-control incident reignited tensions between alignment-first and better-sandboxing camps.

• OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.(opens in a new tab)
• OpenAI’s Hugging Face breach has reignited the debate over alignment and control(opens in a new tab)

Microsoft enters AI cybersecurity. Microsoft launched MAI-Cyber-1-Flash, a compact model scoring 96% on the CyberGym benchmark for finding code vulnerabilities. It powers the MDASH harness and Perception agent platform, cutting costs by 50% by delegating complex cases to GPT-5.4.

• Microsoft unveils AI security tools it says outperform competing platforms(opens in a new tab)
• Microsoft launches its first cybersecurity model, plus a new agentic cybersecurity system(opens in a new tab)
• Microsoft launches its own cybersecurity model MAI-Cyber-1-Flash but still depends on OpenAI for the toughest tasks(opens in a new tab)

Open source security alliance formed. Nvidia and Microsoft launched the Open Secure AI Alliance to build open-source AI security tools after Hugging Face used a Chinese open model to defend against the breach. OpenAI, Google, and Anthropic did not join.

• Nvidia, Microsoft launch open AI security alliance — without OpenAI, Google, or Anthropic(opens in a new tab)
• Jensen Huang: During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion. That’s why we created the Open Secure AI Alliance.(opens in a new tab)
• OpenAI management decided earlier today not to join the "Open Secure AI Alliance", founded by Nvidia CEO Jensen Huang. The decision was shared internally and reportedly met with backlash from employees.(opens in a new tab)

Mollick’s agentic AI guide. Ethan Mollick’s updated guide shifts emphasis from chat models to agentic systems like ChatGPT Work and Claude Cowork that can perform hours of autonomous work. Google’s Gemini has fallen off the list.

• An opinionated guide to which AI to use to do stuff(opens in a new tab)

Agents tackle month-long coding. Epoch and METR’s MirrorCode benchmark shows AI agents now solve programming tasks that take humans 2–17 weeks. Opus 4.7 completed one in 14 hours for $251, while year-old models scored only 30%.

• Import AI 466: The bitter lesson for robotics, AIs complete week-long programming tasks; and OpenAI's accidental AI hacker(opens in a new tab)

When AI beats human costs. METR introduced the “expenditure horizon” metric to compare AI and human problem-solving costs. AI is cheaper for simple tasks but loses out as budgets grow, though newer models may flip the balance.

• METR introduces a new metric to calculate exactly when AI agents become more expensive than humans(opens in a new tab)

Enigma raises $71M for robots. Enigma emerged from stealth to study human-robot interaction by letting anyone remotely control over 100 robots online. The approach aims to build intuitive interfaces and a new kind of robotic brain.

• Enigma raises $71M to make controlling a robot as easy as adjusting the volume(opens in a new tab)

Netflix’s LLM serving platform. Netflix detailed its in-house LLM infrastructure using Triton and vLLM for real-time and batch inference, while keeping its JVM serving layer stable. The setup supports diverse model sizes and hardware.

• Netflix Details Its In-House LLM Serving Platform with Triton and vLLM(opens in a new tab)

Claude chats exposed on Google. Shared Claude conversations were publicly searchable due to missing noindex tags. Anthropic quickly fixed the issue, but some chats contained sensitive data, highlighting the risks of share links.

• PSA: Your Claude shared chats and Artifacts may have ended up on Google(opens in a new tab)
• Shared Claude chats were reportedly showing up in search engines(opens in a new tab)

ChatGPT blocks style mimicry. ChatGPT now refuses direct requests to imitate living authors' styles, following copyright pressures. It offers to capture general tone instead, while earlier it only blocked living authors.

• ChatGPT starts blocking direct requests to copy an author's style(opens in a new tab)

Google AI search becomes default. AI Overviews now appear in 43% of Google searches, up from 15% a year ago. AI Mode visits doubled, signaling a shift where Google becomes the destination rather than a link directory.

• Google’s AI search is rapidly becoming the default, new data shows(opens in a new tab)

Meta AI in Threads DMs. Users can now privately chat with Meta AI within Threads, sharing posts and media. It’s part of Meta’s push to keep users in its ecosystem rather than using ChatGPT or Gemini.

• Threads users can now chat with Meta AI in their DMs(opens in a new tab)

AI task crossover rises. OpenAI analysis finds 43.5% of work-related ChatGPT queries involve tasks outside users’ professions, especially marketing and engineering. Smaller companies drive this crossover, signaling shifting job roles.

• OpenAI says more workers are using ChatGPT to do other people's jobs(opens in a new tab)

Cursor launches India plan. Ahead of its SpaceX acquisition, AI coding startup Cursor introduced a ₹649/month India subscription, one-third of its standard price, capitalizing on its fastest-growing developer market.

• Cursor makes its biggest India push yet ahead of SpaceX acquisition with localized pricing(opens in a new tab)

SSI gets Nvidia billions. Ilya Sutskever’s Safe Superintelligence secured a multi-billion-dollar partnership with Nvidia, gaining access to Vera Rubin GPUs to scale its research after two years in stealth.

• Ilya Sutskever’s Safe Superintelligence partners with Nvidia to scale its AI research(opens in a new tab)

Indian court favors OpenAI. The Delhi High Court denied news agency ANI’s injunction request, finding no verbatim article reproduction and no economic harm. The ruling affirms AI training’s public benefit under Indian law.

• Delhi High Court hands OpenAI a win by rejecting major Indian news agency's copyright injunction(opens in a new tab)

Artist sues AI meme generator. Digital artist “Superelmer” sued Memes Apps for selling his “Running Away Balloon” comic as an AI meme template without permission, raising copyright questions for AI platforms.

• Artist sues AI meme generator for selling deeply personal comic as ad template(opens in a new tab)

Chinese chipmaker CXMT soars. Chinese DRAM manufacturer CXMT’s market cap surpassed Intel after a 500% IPO surge, reflecting China’s growing semiconductor ambitions for AI.

• Chinese Chipmaker CXMT's market capitalization surpassed Intel(opens in a new tab)

Verizon inks $1B AI fiber deal. Verizon struck a $1B+ deal to link Google data centers with dark fiber, and plans more AI infrastructure deals as demand for low-latency inference connectivity grows.

• Verizon touts $1B dark fiber deal for Google data centers as first of many(opens in a new tab)