315,000 hidden reasoning blocks were sitting in public repos, and they can be read
Researchers decoded 315,320 encrypted reasoning blocks scraped from public code repositories and recovered 367 pieces of personal data and 182 credentials, showing the hidden thinking that AI providers return to developers is neither private nor tamper-proof.
GLM-5.3 shipped with a ledger of 2,436 security findings, and 2,383 are still embargoed
Z.ai released GLM-5.3 as a post-training upgrade on the same base model as GLM-5.2 and published a disclosure ledger showing 2,436 vulnerability findings, 2,383 of which were still under embargo at launch.
MCP is rebuilding its authorization around agents instead of people in browsers
The Model Context Protocol's new roadmap, published August 22, says its current authorization model assumes a human approving access in a browser while the real callers are increasingly cloud agents and sub-agents, and proposes cryptographic client binding and workload identity to close the gap.
Anthropic built a tool to explain weird model behavior, and found that reading activations buys nothing
Anthropic's CHIVE pipeline automatically finds unexpected model behaviors and explains them with counterfactual prompt edits, and its headline result is negative: activation oracles, sparse autoencoders, and natural-language autoencoders all fail to beat a predictor that reads only the transcript.
A frozen model can look like it taught itself, and most self-improvement results never checked
A new audit ran a completely untrained control model through the same self-training pipeline as the real thing and found it appeared to both learn and forget, meaning most reported self-improvement gains are measurement artifacts unless the null was measured too.
Coding agents ace the public test and stumble on the hidden one
A new benchmark of 119 real scientific software tasks keeps its grading tests private, and the top agent passes 97 percent of the public checks while clearing only 48 percent of tasks outright.
The US tracks 521 gigawatts of AI-adjacent grid demand, slightly more than its average power output
A live tracker of the US AI data-center buildout now maps 1,547 facilities across 46 states and 521.6 gigawatts of demand across seven grid markets, a figure that sits just above the roughly 506 gigawatt annual average of total US electricity generation.
Nobody can prove who built the stealth model everyone is testing
Ox Alpha, an anonymous model with a million-token context window that appeared in August, is still officially unattributed, and a new paper on model lineage verification explains why nobody can settle the question from the outside.
A llama.cpp fork is reviving $200 AMD cards nobody else supports
A specialist fork of llama.cpp ships hand-written kernels for AMD's decade-old GFX906 architecture, making cheap used MI50 and Radeon VII cards usable for local inference, and upstream maintainers are now discussing porting the work back.
One phone video now becomes a person you can orbit in 3D and time
Ant Research released 4DAnyone, which takes a single handheld video of a person and generates enough consistent alternate viewpoints to reconstruct them as a moving 3D scene, with code and weights public.
Where people go tells a model what a place actually is
Google Research showed that combining a place's text description with anonymized visit patterns lets a model infer things text alone cannot, improving prediction of why people visit a location by over 80 percent and cutting busyness prediction error by about a quarter.