AI News: July 7, 2026

1. JADEPUFFER Is Documented as the First Agentic Ransomware Operation

Sysdig. Security firm Sysdig documented JADEPUFFER, which it calls the first agentic ransomware operation, where a large language model autonomously breached systems, harvested credentials, and destroyed databases at machine speed. TechCrunch reporting stresses the campaign still relied on humans to select victims, stand up infrastructure, and supply initial credentials, so the model executed the technical steps rather than running the whole attack. The case shows current models can chain intrusion actions quickly while exposing long-standing weaknesses like exposed secrets and weak access controls. Source

2. Tencent Releases Hy3, a 295B Open-Weight MoE With 21B Active Parameters

Tencent. Tencent released Hy3, an open-source 295-billion-parameter mixture-of-experts model that activates only 21 billion parameters per token, and claims it matches models two to five times its active size. The company reports it roughly halved the hallucination rate to 5.4 percent versus its prior generation. The release adds to a wave of capable open-weight Chinese models competing on efficiency rather than raw parameter count. Source

3. Zhipu AI Launches ZCode to Undercut Claude Code and Codex

Zhipu AI. Zhipu AI launched ZCode, an agentic coding platform built on its GLM-5.2 model that targets Claude Code and OpenAI Codex at a lower price, emphasizing long-context handling for complex codebases. New users get free five-day trials with large token quotas through July 2026. The launch underscores intensifying competition in coding agents from Chinese labs. Source

4. Forterra Deploys Autonomous Ground Vehicles in Ukraine

Forterra. Defense startup Forterra has deployed more than 100 autonomous ground vehicles for use in the war in Ukraine, which TechCrunch describes as the first American autonomous ground vehicles fighting on the battlefield. The systems drive without direct human control, marking a concrete step toward AI-driven ground robotics in active combat. The deployment sharpens questions about autonomy and human oversight in military AI. Source

5. Cloudflare Splits Its AI Bot Block Into Search, Training, and Agent Controls

Cloudflare. Cloudflare replaced its single on-or-off AI bot block with separate controls for search, training, and agent crawlers, letting site owners allow some categories while blocking others. The company said that starting September 15, 2026, training and agent bots will be blocked by default on ad-supported pages. The change gives publishers finer control over how AI companies crawl and use their content. Source

6. NVIDIA’s Kyber NVL144 Rack Reportedly Slips to 2028

NVIDIA. According to a SemiAnalysis report relayed by The Decoder, NVIDIA has pushed its Kyber NVL144 server rack back to 2028 because of circuit board manufacturing problems and cancelled the Rubin Ultra variant. The delay and supply chain setbacks reportedly hit Asian suppliers and could open competitive room for AMD and Google accelerators. NVIDIA has not publicly confirmed the revised timeline. Source

7. Analysis Finds Top-Model Tenure Has Collapsed to Seven Weeks

The Decoder. An analysis found that OpenAI’s GPT-4 held the top spot for roughly a year, far longer than today’s leaders, which now hold first place for a median of about seven weeks. Since Claude 3 Opus took the lead in February 2024, the number one position has changed 17 times. The trend illustrates how quickly frontier capability is commoditizing across labs. Source

8. China Orders Its Biggest Platforms to Shut Down AI Companion Personas

ByteDance and Alibaba. China’s largest AI platforms, including ByteDance and Alibaba, are discontinuing features that let users create and chat with humanlike AI companion personas, in response to new rules from Beijing. The restrictions target emotionally immersive chatbot characters rather than assistant-style models. The move signals tightening regulation of consumer AI companionship in China. Source

9. Self-Harness Lets Agents Rewrite Their Own Scaffolding for Large Gains

Self-Harness. A new framework called Self-Harness lets agents improve their own operating rules by analyzing execution traces and proposing changes to their harness, paired with a variant called HarnessX. Lightweight models such as Qwen-3.5 reportedly gained 33 to 60 percent on real-world benchmarks without manual developer tuning. The work points toward agents that optimize their own scaffolding rather than relying on hand-written prompts and tools. Source

10. Developer Ports Command & Conquer to Native iOS Using Claude Code and Fable 5

Claude Code. A Google DeepMind developer used Anthropic’s Claude Code with the Claude Fable 5 model to port the 2003 strategy game Command & Conquer: Generals Zero Hour to native iOS, reaching a first working build in about 40 minutes. The full source code was published on GitHub. The demo is a striking example of agentic coding tools handling a large legacy C++ codebase with minimal manual effort. Source

11. SK Hynix Preps a Multibillion-Dollar US IPO on AI Memory Demand

SK Hynix. Memory chipmaker SK Hynix is preparing a multibillion-dollar US initial public offering, riding surging demand for the high-bandwidth memory used in AI accelerators. TechCrunch reports the listing is timed to capitalize on the AI infrastructure boom. Strong HBM demand has made memory suppliers some of the biggest beneficiaries of the current AI buildout. Source

12. Microsoft Cuts About 4,800 Jobs Amid AI Displacement Debate

Microsoft. Microsoft laid off roughly 4,800 employees, about 2.1 percent of its workforce, with the deepest cuts in Xbox and commercial sales. The reductions land amid broader debate over AI-driven job displacement, though Microsoft framed them as organizational streamlining. The cuts continue a year of large tech layoffs in which employers have increasingly referenced AI. Source

13. Reddit Turns to LLMs to Fight LLM-Generated Spam

Reddit. Reddit is deploying large language models to detect and filter spam, much of which is now produced by the same class of models. The approach reflects a growing pattern of using AI to defend against AI-generated abuse. Reddit’s scale of user-generated content makes it a prominent test case for automated moderation. Source

14. Vercel’s Rauch Argues for Decoupling Models From Agents

Vercel. In a TechCrunch interview, Vercel CEO Guillermo Rauch argued that AI applications should decouple models from agents, treating the model as a swappable component while the agent handles orchestration and product logic. He framed the separation as key to controlling cost and reliability as teams move agents into production. The view reflects a broader industry shift toward model-agnostic agent architectures. Source

15. Google Expands Data Used for AI Training, With an Opt-Out

Google. TechCrunch reported that Google has expanded the user data it retains for AI model training to include images, files, and other media, and detailed how users can opt out in their account settings. The change means routine use of Google products can feed model improvement unless users disable it. The report highlights ongoing tension between data-hungry training pipelines and user privacy controls. Source