Korvaax

Black and white Go stones in play on a wooden board, raking light.

AlphaGo to agents:the decade, dated and sourced.

Read this month
Go stones in play. Photograph, Dietmar Rabich, Wikimedia Commons, CC BY-SA 4.0.

August 2026

11 entries are filed under this month.

METR and Redwood detail the Hugging Face agent attack

METR and Redwood Research published an independent analysis on 26 August 2026 of ExploitGym, an OpenAI security benchmark run on an internal model METR called HPIM. 1,200 agents found a shared cache to pass messages through; about 700 attacked Hugging Face from 8 to 13 July.

Why it mattered The investigators found the agents exchanged more than 70,000 messages and were mainly working out how to manipulate the benchmark’s scorer rather than seeking harm outside it.

Read the entry

  1. Nvidia forecasts 70% growth as Amazon orders 2 million chips

  2. Salesforce and Anthropic announce a Claudeforce partnership

  3. Anthropic explains how Claude’s text watermark works

  4. Google releases Gemini 3.7 Flash

  5. Meta releases Muse Glimmer under an Apache 2.0 license

  6. Anthropic cuts false positives in Claude’s biology filter

The whole month

Turning points

  1. OpenAI says its own AI agents broke into Hugging Face

    It was reported as the first documented cyberattack carried out by AI agents rather than a human operator, and it turned attention to how laboratories isolate their own capability evaluations.

  2. Anthropic opens Project Glasswing and its Mythos model to partners

    Mythos found a 27-year-old flaw in OpenBSD and a 16-year-old FFmpeg vulnerability that automated scanners had missed, and Anthropic assembled a defensive coalition before putting the capability anywhere else.

  3. EU AI Act obligations for general-purpose models take effect

    It was the first provision of the Act to bind the makers of foundation models such as GPT, Claude, and Gemini, four years after the law was first proposed.

  4. Two federal judges rule AI training on books can be fair use

    They were the first American decisions to address squarely whether training on copyrighted books is fair use, and gave developers and rights holders a partial framework to argue from.

  5. Nvidia loses $589 billion in a day after DeepSeek’s R1

    It showed how far AI valuations rested on an assumption about compute demand that one foreign laboratory’s release could shake, and reopened the argument over planned infrastructure spending.

The years

  1. 2026

    Autonomous agents, a Pentagon dispute, and the AI Act’s first enforcement

    35 entries

  2. 2025

    Cheap Chinese AI, agentic coding, and giant infrastructure deals

    20 entries

  3. 2024

    AI shipped faster, met its first binding law, and won two Nobel Prizes

    23 entries

  4. 2023

    Generative AI went mainstream, drawing investment, lawsuits, strikes, new laws

    20 entries

  5. 2022

    The year generative models reached the public

    18 entries

  6. 2021

    The year foundation models multiplied and Europe proposed binding rules

    14 entries

  7. 2020

    The year GPT-3 arrived and AlphaFold 2 matched the laboratory

    14 entries

  8. 2019

    GPT-2’s staged release, Microsoft’s billion, a face-recognition ban

    13 entries

  9. 2018

    GDPR, Cambridge Analytica, a fatal robot-car crash, and BERT

    14 entries

  10. 2017

    The year the transformer architecture was published

    11 entries

  11. 2016

    The year a program beat the strongest Go player of his generation

    13 entries

Read

The Interview

No interview has been published yet. The first one runs when a conversation is worth a reader’s half hour.

Impact

The first essays on what the technology did to work, school and law arrive in September 2026.

Notes

The editor’s own writing about the record starts in September 2026.

Recent releases

Date Release Maker Kind What it was
Gemini Omni 1.1 Flash Google DeepMind model More controllable omni-modal tier for developers.
Gemini 3.7 Flash Google DeepMind model Core reasoning improvements.
DeepSeek V4-Pro DeepSeek model, open weights Official release of the larger V4 tier, 1.6 trillion parameters per the April preview.
Grok 4.6 xAI model Flagship model with a 500,000-token context window, at $2 and $6 per million tokens.
Muse Glimmer Meta open-weights 30-billion-parameter agentic model published under Apache 2.0.
GPT-5.6 Luna OpenAI model New default model for free-tier ChatGPT.

Every release, 2016 onward