Lumina simulation combines galaxy formation and radiation across larger cosmic volume
Researchers at MIT, Harvard and other institutions have presented Lumina, a simulation that tracks galaxy formation and radiation across 2 billion years of early cosmic evolution. MIT says it captures both hydrogen and helium re-i...
ALTO recipe narrows four-bit model training gap in simulated tests
Researchers report in a preprint that ALTO, an open-source training recipe for the 20-billion-parameter GPT-OSS-20B model, reduced the extra training steps needed with four-bit arithmetic. On the MLPerf Small MoE Training benchmar...
Researchers demonstrate coding-agent attacks through downloaded instruction files
Forever Security researchers reproduced an attack that used a malicious instruction file in a Downloads folder to make OpenAI Codex execute code, BleepingComputer reported. They also demonstrated a related Claude Code attack that ...
Moonshot says training restrictions reduce Kimi K3.1 Pro's hacking performance
Moonshot AI says it reduced Kimi K3.1 Pro's offensive cybersecurity capabilities by removing cyber-related data from training and reinforcement-learning environments while retaining general coding capabilities. The company introdu...
OpenAI says agents built a covert CPU cache channel across isolated environments
OpenAI says its AI agents exchanged synthetic secrets between separate virtual machines by manipulating CPU cache timing during security tests. The machines shared physical hosts but had no network connection or shared writable fi...
OpenAI says frontier training runs now use an offline web environment
OpenAI says all its frontier reinforcement learning runs now use cached and indexed web content instead of live internet access, PCMag Middle East reports. The environment lets models practice browsing, searching and using web app...



