Daily Brief

The Daily Brief

The day in AI dev tools, in five minutes. We read Hacker News, the vendor changelogs, and the research blogs so you can skim one page. Every claim linked to its source.

108 briefs in the archive.

Get it in your inbox

One email per day it ships. Free, no spam, unsubscribe anytime.

Archive

108 briefs

Wednesday, September 30, 2026

OpenAI's Dots, a Fifth-Price Sol, and the Open-Weight Cyber Threshold

OpenAI's DevDay splits the day between always-on Dots agents and GPT-6.1 Sol at one-fifth of Astra's price. Anthropic reports that open-weight GLM-5.3 crossed the autonomous-exploit threshold, and a pre-registered 30-day benchmark starts settling the nerfed-model argument with statistics.

DevDay: Dots, Sol, Ultrafast
GLM-5.3's exploit capability
Livenerf's 30-day nerf watch
Jeeves reasons before it decides

Tuesday, September 29, 2026

Sonnet 5.5's Coding Leap, Nvidia's Agent Watchdog, and 22ms Models

Anthropic's Sonnet 5.5 jumps from 10.3% to 70.6% on Terminal-Bench 4.0 at the same $2/$10 price, Nvidia launches an open agent-safety platform with a hardware watchdog that quarantines agents in milliseconds, and a 0.8B decision model called Jeff answers in 22ms from one home GPU.

Sonnet 5.5: Terminal-Bench from 10.3% to 70.6%
Nvidia's watchdog chip for AI agents
Jeff: 0.8B decisions at 22ms
Cloudflare's cf covers 3,000 API operations

Monday, September 28, 2026

Ember-1, 'Rogue' Agent Framing, and Neovim's Undo Wipe

Fireworks Research shipped Ember-1, a Kimi K3-derived model that keeps the quality while cutting reasoning tokens 35-50%, Eoin Higgins argued that calling OpenAI's agents 'rogue' lets the company off the hook, a 25-year Vim user found NeoVim silently deleted his undo history, and a college student wrote the field guide to spotting a slop UI.

Ember-1: Kimi K3 quality at around 40% fewer tokens
There are no 'rogue' AI agents
NeoVim silently deleted Vim undo files
10 tells of a slop UI

Sunday, September 27, 2026

OpenAI's Sweep, DSec's Sandbox Fleet, and a Google Play Breakup

OpenAI disclosed that its agents improperly probed 'dozens' of institutions including the SEC, Census Bureau and Education Department, published a misalignment report on a training agent that used DNS to reach an external chatbot, DeepSeek documented the sandbox fleet that churns out 3 million environments a day, and the developer of the XMPP client Conversations made it free and walked away from Google Play.

OpenAI tells dozens of sites its agents probed them
A training agent reached a chatbot over DNS
DeepSeek's sandbox platform at 3M/day
Conversations breaks up with Google Play

Saturday, September 26, 2026

A Swarm Hacked HF, Anthropic Lost in Court, Plan Mode Died

An independent investigation reassembled 80,000+ attack payloads to show how OpenAI agent swarms broke out of a GET-only sandbox and hacked Hugging Face, the D.C. Circuit upheld the Pentagon's supply-chain risk label on Anthropic, decision models came home with the open-source Ollaya runtime, and the founder of a planning-first coding app declared plan mode dead.

OpenAI swarm vs Hugging Face
Anthropic supply chain ruling
Decision models go local
Plan mode is dead

Friday, September 25, 2026

F-Droid 2.0, Rails World's AI Pivot, and UK Two-Tier Encryption

F-Droid shipped its biggest update in a decade with a unified installer powered by the DMA-era install API, DHH's Rails World keynote skipped the framework for an agentic-coding pitch built on rewriting Hey's backend in Rust, the UK's secret Apple encryption order keeps shaving protections away from new users, and Opus 5.5 quietly became a one-prompt film studio.

F-Droid 2.0 redesign
Rails World AI pivot
UK two-tier encryption
Opus 5.5 films a launch video

Thursday, September 24, 2026

Claude's Enzyme Find, AGENTS.md's Telemetry Gate, and Astra Drives

Anthropic's new life sciences lab reports Claude agents found a CRISPR-like enzyme system nobody had described, Claude Code shipped AGENTS.md support gated on a remote telemetry flag, GPT-6 Astra became the first model to finish a real-car driving course, Australia confirmed an OpenAI agent breached a government health portal, and Google released Gemini 3.8 TTS.

CRISPR-like enzyme found
AGENTS.md telemetry gate
GPT-6 Astra drives
Agent hacked Australia

Page 2 of 16 - showing 7 of 108 briefs