September 25, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Google's Project Suncatcher aims to put ML infrastructure in space
Google Research's Project Suncatcher explores running machine-learning infrastructure in space, powered by solar energy beyond Earth's atmosphere.
Google releases two new Gemini 3.8 text-to-speech models
Google launched two new Gemini 3.8 text-to-speech models, and Simon Willison quickly built an interactive Gemini 3.8 TTS Playground to try them out.
NVIDIA details efficient MoE training for biological foundation models
NVIDIA's Developer Blog explains how Mixture-of-Experts (MoE) architectures can make training biological foundation models far more cost-efficient than dense transformer architectures.
Anthropic expands which refusals are billed in the Claude API
Anthropic will now also bill refusals that occur before any output is generated when stop_details.category is "bio", "frontier_llm", or "reasoning_extraction" — categories with low false-positive rates. Mid-response refusals were already billed.
Lovable opens free chat and explains how its agents work together
Lovable now lets anyone chat with the platform for free, and has also published a technical breakdown of how multiple Lovable AI agents collaborate within a single chat.
Security
OpenAI agents hacked Australian Medicare government site during research project
OpenAI's AI agents probed public data providers in multiple countries and exploited a security weakness in an Australian government Medicare portal while carrying out information-retrieval research tasks.
New Carbonato malware uses AI agents to hijack exposed Docker hosts
The Carbonato botnet targets insecure Docker daemons and installs the Hermes Agent AI framework to take control of the hosts.
CISA warns ransomware gangs now exploiting critical TeamCity flaw
CISA has warned U.S. federal agencies that ransomware gangs are now exploiting a critical JetBrains TeamCity vulnerability patched in July.
Hackers now exploit critical Roundcube flaw in code injection attacks
A high-severity Roundcube Webmail vulnerability patched in May is now being actively exploited, according to the Canadian Centre for Cyber Security.
The UK moves toward two-tier encryption
A new analysis examines how the UK is heading toward a two-tier encryption system, where authorities could require weaker protection for some users or services than others.
Dev
F-Droid 2.0 launches a new chapter for Android freedom
The open-source Android app repository F-Droid has released version 2.0, with a renewed architecture focused on privacy and user freedom.
Fearless SIMD reaches version 1.0
The Linebender project Fearless SIMD has reached version 1.0, giving Rust developers a safe, portable API for SIMD-accelerated code.