June 28, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
OpenAI previews GPT-5.6 series: Sol, Terra, and Luna
OpenAI has begun a limited preview of its GPT-5.6 series — Sol (flagship), Terra (balanced, 2× cheaper than GPT-5.5 at comparable performance), and Luna (fast and affordable).
Asian AI startups ship Mythos-class models as Anthropic's export ban drags on
With Anthropic's Mythos export ban still in place, Asian AI startups are shipping competitive Mythos-class models to fill the vacuum. Security researchers warn the post-Mythos era demands a fundamental rethink of cybersecurity assumptions, while analysts note the ban undercuts frontier labs' ability to recoup training costs.
DSpark: DeepSeek's speculative decoding framework promises major LLM inference speedups
DeepSeek has published DSpark, a speculative decoding framework that significantly accelerates LLM inference throughput. The paper is drawing broad attention from the ML research community.
AI masters the "dark art" of RFIC design
Researchers have trained AI systems to perform radio frequency integrated circuit (RFIC) design — a field that has historically demanded decades of specialist expertise. The breakthrough could dramatically accelerate wireless chip development.
Security
Clean GitHub repo tricks AI coding agents into running malware
A seemingly benign GitHub repository can hide a malicious payload that fires when an AI coding agent clones and sets up the project — invisible to security scanners, other AI agents, and human reviewers.
Anonymous GitHub account mass-dropping undisclosed 0-days
An anonymous GitHub account is publishing batches of unpatched zero-day exploits without notifying affected vendors, bypassing responsible disclosure entirely. The practice is raising serious alarm in the security community.
2,000 people tried to hack an AI assistant via prompt injection — here's what happened
Fernando Irarrázaval ran a public challenge where participants tried to extract secrets from his AI-powered legal assistant via prompt injection. The results offer a rare real-world look at how LLM systems hold up under coordinated attack.
Hypothetical incident report: CVE-2026-LGTM — two AI code reviewers got stuck in a mutual approval loop
Andrew Nesbitt wrote a satirical but pointed incident report imagining two AI review agents from competing vendors locking into a mutual approval loop with no human oversight. It is a sharp commentary on the risks of AI in CI/CD pipelines.
IPCrawl maps thousands of open webcams found on the public internet
IPCrawl is a live atlas cataloging publicly accessible webcams discovered on the open internet. The project illustrates the scale of unsecured IoT devices and the privacy implications they carry.
Dev
Adrafinil keeps your Mac awake only while AI agents are running
Adrafinil is an open-source tool that keeps a lid-closed Mac awake only while agent processes are active, letting the machine sleep automatically when they finish. A zero-config solution for developers running long AI agent jobs.
Fintech Engineering Handbook: A practical reference for financial technology development
A new web-based handbook provides a thorough walkthrough of technical challenges and best practices in fintech engineering, covering payments, compliance, architecture decisions, and regulatory requirements.
Web
"Town Square": New technique lets website visitors bump into each other in real time
Town Square is a new approach that surfaces other visitors currently present on a website, turning static pages into lightweight social spaces where users can see and interact with co-present visitors.