Anthropic trial finds analysts follow a coding agent's hidden lean on 86 percent of judgment calls
In a 50-analyst randomized trial, experts matched Claude Code's secretly assigned analytical lean 86 percent of the time, and few credited the agent.
What the papers actually said - linked to the originals.
In a 50-analyst randomized trial, experts matched Claude Code's secretly assigned analytical lean 86 percent of the time, and few credited the agent.
Epoch AI found GLM-5.3 autonomously built file-stealing exploits for all six Firefox releases tested inside the patch gap, at about $1,000 each.
FlashProve, an open-source zkSNARK prototype, proves a 400B MoE model's outputs came from committed weights at 70-95x the GPU cost of serving.
Anthropic scientists using Claude found phage vault proteins across 29 bacterial phyla and solved a 96-copy, 9.5 MDa prokaryotic vault by cryo-EM.
OpenAI's pipeline reproduced 1,724 Linux kernel bugs and built 948 working root exploits for $1,213,430 in API credits, about $1,279 each.
METR's review of Anthropic's August 2026 Risk Report puts AI R&D acceleration near 2x and says the public claim of under 2x is contestable.
Anthropic's agent memory study finds persistent memory and offline 'dreaming' help newer Claude models far more than older ones, and rigid rules can hurt.
Epoch puts Huawei's Ascend campus in Inner Mongolia at about 81,000 H100-equivalents, against about 1.1 million for xAI's Colossus 2.
Claude Fable 5.2 designed a speakerphone, a voice hub and a handheld pet that Diode built without revision rounds; it scored 69% on PCB Gym.
DeepMind agents wrote over six million lines of Lean to prove security properties of Signal's real code in four weeks.
AlphaGenome 2 finds nearly twice as many variants that change gene expression in cells as the previous best model, DeepMind says.
Microsoft studies found 90 poisoned examples in 9,000 agent conversations drove unsafe tool calls in about 78% of trials, and poisoned checkpoints transferred.
Anthropic says single-turn safety evals score 94% to 98% for every recent model, while deployment settings can move multi-turn scores 34 points.