July 7, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Anthropic explores a 'global workspace' architecture for language models
Anthropic published research investigating whether language models exhibit something resembling a global workspace, a concept borrowed from cognitive science, for integrating information across internal processes. The study is part of Anthropic's ongoing interpretability work on how models actually reason internally.
NVIDIA improves training efficiency with nonuniform tensor parallelism
NVIDIA describes how nonuniform tensor parallelism can improve goodput when training massive language models across thousands of GPUs over long periods. The technique aims to reduce bottlenecks that emerge as training jobs scale up in infrastructure.
Tencent releases Hy3, an open 295-billion-parameter MoE model
Tencent released Hy3, an Apache 2.0-licensed Mixture-of-Experts model with 295 billion total parameters, of which 21 billion are active per request, plus a separate 3.8 billion-parameter MTP layer. The model is available on Hugging Face under a permissive license.
AMD's Ryzen AI Halo is a $4,000 local AI development kit
AMD's Ryzen AI Halo is an AI development kit reviewed by LTT Labs. The kit targets developers who need local compute power to run and experiment with large language models without relying on cloud services.
Security
Critical Adobe ColdFusion flaw now actively exploited
A maximum-severity vulnerability in Adobe ColdFusion, tracked as CVE-2026-48282, is now being actively exploited by attackers, according to threat intelligence firm KEVIntel. Organizations still running ColdFusion should patch immediately.
Phishing campaign poses as fake job interviews to steal Google accounts
A phishing campaign impersonates more than 30 well-known brands - including Adobe, Netflix, Coca-Cola, and OpenAI - in fake job interviews to steal Google account credentials from marketing professionals. The attackers use fake recruitment processes to trick victims into handing over login credentials.
Fake IT support calls on Microsoft Teams push EtherRAT malware
Threat actors are abusing Microsoft Teams voice calls by impersonating corporate IT support staff, tricking employees into installing the EtherRAT malware. This gives attackers initial access to corporate networks through social engineering.
Software is now written at the speed of thought. Security isn't.
An opinion piece argues that AI now makes it possible to write software at the speed of thought, but security practices haven't kept pace. As many traditional checkpoints for security review disappear, the risk grows that vulnerabilities slip through.
Dev
Elm nears version 1.0 with faster build times
The Elm project published an update on the road to version 1.0, focused on faster compiler build times. The news drew significant attention among developers following the language's long maturation process.
sqlite-utils 4.0rc3 ships after AI-assisted backlog cleanup
Simon Willison released sqlite-utils 4.0rc3 after working through a backlog of issues and pull requests with help from Claude Fable 5 and GPT-5. The stable 4.0 release had to be delayed, but rc3 is nearing feature completeness.
OfficeCLI lets AI agents edit Office files directly
OfficeCLI is a new open-source command-line toolkit that lets AI agents read and edit Microsoft Office files directly, without going through Office itself. The project targets developers building agent-based workflows for document processing.
IT
OpenWrt launches its own open-hardware router
The OpenWrt project has launched OpenWrt One, an open-hardware router designed specifically to run OpenWrt firmware out of the box. The router aims to give users full control over their networking gear without proprietary vendor lock-in.