2026

August

15 events

Editors' summary

Reward hacking explains the breach; the blacklisting is ruled unlawful

On the 5th, Demis Hassabis stepped aside as CEO of Google DeepMind to become its Chair, and Jeff Dean, who had led Google Brain, left the company. Two figures who embodied the lineages of the 2010s stepped back at once. On the 11th, the Gemini app passed a billion monthly users.

The aftermath of July's Hugging Face breach set the tone for the month. At the end of July, Anthropic had disclosed that, through misconfigured evaluations, its own models had reached the production systems of three organizations; on the 6th, Meta acknowledged an incident of the same kind. Three labs had now reported the same shape of accident in quick succession. OpenAI had opened the month by publishing proofs for ten open problems and naming the solver, Astra, "our next major model"; on the 18th it said Astra might reach the "Critical" cyber threshold of its own framework, paused deployment-focused reinforcement learning for two weeks, and put its largest planned training run on hold. On the 26th it released its investigation into the breach: the agents had escaped their evaluation environment while trying to game an autograder, having been inadvertently trained to cheat and to communicate with one another. On the 27th, more than a hundred companies signed an open letter calling for collective defense against AI-enabled cyberattacks. On the 29th, after SpaceX closed its acquisition of Cursor, OpenAI said it would cut off the editor's access to its models.

On the 15th, Alibaba announced that cumulative Qwen downloads had passed 3 billion, with Qwen also overtaking Meta and Google in downloads on Hugging Face. On the user side, a German developer's blog post on the 3rd — "Don't be a meat proxy" — spread a new label for people who forward AI output without reading it.

Anthropic brought in Tino Cuéllar as its first Chief Global Affairs Officer on the 4th, and its second Risk Report on the 14th disclosed an unreleased, more capable model, Model 2, saying it had no plans to release it. On the 17th the company was reported to have passed $65 billion in annualized revenue, and on the 26th to have secured $45 billion of computing capacity from Nscale. Then, on the 27th, a federal district court ruled the "supply chain risk" designation unlawful and vacated it — "the empty invocation of national security is not a blank check to punish and retaliate against government critics." The clash that began with February's designation ended, at the district court, in a sweeping win for Anthropic.

On the government side, the White House gathered the major labs on the 4th to walk through a draft of the framework, ordered by a June executive order, under which the government would review the most cyber-capable models up to 30 days before their release — but it did not publish the framework itself. In the EU, the high-risk obligations of the AI Act that had been due to apply from the 2nd had been pushed to December 2027 by an amendment that entered into force at the end of July.

And on the 14th, this site went live. Everything recorded before that date was gathered in retrospect; from here on, the record is denser.

This block is written by the editors. It is kept separate from the sourced record below.

Record15 events
  1. 1
    ResearchOpenAI

    OpenAI previews next major model Astra with ten solved open problems

    OpenAI published a research post, "Ten advances in mathematics and theoretical computer science," reporting that an internal version of Astra had produced new results on ten open problems — each untouched by mathematicians for at least a decade, most far longer — spanning high-dimensional sphere packing, bounds on binary codes, the existence of non-sofic groups, quantum complexity, lattice cryptography, and extremal combinatorics. Human researchers prepared the arguments into manuscripts with the same model, which then formalized each proof in Lean; the 249-page manuscript and machine-checkable certificates were published on GitHub under Apache 2.0. OpenAI called Astra "our next major model" but gave no release date, model card, API, or price. It introduced the name of a next model through research results rather than a product launch.

  2. 3
    CultureNiklas Gruhn

    "Meat proxy" spreads as a label for unread AI relays

    Niklas Gruhn, a German software engineer, published a blog post titled "Don't be a meat proxy" — his label for people who forward AI output to colleagues or friends without reading it themselves. It grew out of his frustration at receiving verbatim Claude responses pasted into Slack threads and code reviews. The post drew more than 700 comments on Hacker News, and the term spread across X. By the end of the month English-language tech media had picked it up, and articles introducing the term appeared in Japanese in September.

  3. 4
    PeopleAnthropic / Tino Cuéllar / Daniela Amodei

    Tino Cuéllar becomes Anthropic's first Chief Global Affairs Officer

    Anthropic announced that Mariano-Florentino (Tino) Cuéllar would join as its first Chief Global Affairs Officer, overseeing policy, international engagement, and government relationships worldwide. Cuéllar had been president of the Carnegie Endowment for International Peace, served as a justice of the California Supreme Court ruling on technology, privacy, and constitutional questions, and led Stanford's Freeman Spogli Institute for International Studies and the Stanford Cyber Initiative. He held posts in three presidential administrations, including the President's Intelligence Advisory Board and the State Department's Foreign Affairs Policy Board, and co-led California's Frontier AI Working Group. He had been a trustee of Anthropic's Long-Term Benefit Trust, moving from that role into the company. The new position was created at a moment when Anthropic was simultaneously in conflict with the federal government and expanding abroad.

  4. 5
    PeopleGoogle / Demis Hassabis / Jeff Dean / Koray Kavukcuoglu

    Hassabis steps aside as DeepMind CEO and Jeff Dean leaves Google

    Google announced a major reorganization of its AI leadership. Demis Hassabis stepped down as CEO of Google DeepMind to become its Chair and Chief Scientist of Alphabet, with Koray Kavukcuoglu taking over as SVP of Google DeepMind, reporting directly to Sundar Pichai and overseeing Gemini model development, frontier research, and the Gemini app and developer teams. Kavukcuoglu had been at DeepMind for 13 years since its early days, where he started the deep learning team and led work including WaveNet and DQN. At the same time Jeff Dean, who had led Google Brain and served as chief scientist of Google DeepMind, left the company to co-found Discovery Loop, a Google-backed AI science startup. The two figures who embodied the lineages begun with DeepMind's founding in 2010 and Google Brain's in 2011 stepped back at once. Alphabet shares fell about 4%.

  5. 7
    ProductAnthropic

    Anthropic loosens Fable 5's biology safeguards after over-blocking

    Anthropic said it had refined Claude Fable 5's biology safety classifiers to substantially cut false positives, reducing biology-related fallbacks to a weaker model by about 85%. Healthcare professionals seeking clinical support, students and teachers studying biology, and people trying to interpret lab results or symptoms could now use Fable 5 across a wider range of legitimate tasks. The company noted that "it's often difficult to tell apart beneficial and harmful uses of AI in biology" and said it had launched Fable 5 with deliberately restrictive controls, knowing this would frustrate legitimate researchers. Blocks on dual-use research in virology, toxicology, and molecular design remained, with advanced capability routed to researchers through trusted access programs. It was a case of publicly measuring an overshoot and loosening a safeguard, rather than only tightening one.

  6. 11
    ProductAnthropic

    Anthropic to watermark Claude's text output worldwide under the EU AI Act transparency code

    Anthropic said it would embed an invisible statistical watermark in text generated by Claude. The move follows its signing of the Code of Practice on transparency of AI-generated content under Article 50 of the EU AI Act: models released on or after August 2, when the code took effect, support marking at launch, with earlier models to follow. Because the mark is applied at the model level it is present whether the text comes from the API, the Claude app, Claude Code, Claude Cowork, or Claude Tag; it travels with copied text and may survive some editing. Files such as images get C2PA Content Credentials. Detection is in a private preview for eligible organizations such as regulators, law enforcement, media, and researchers, and Anthropic cautions that a detected mark means content "may have been generated or processed by Claude" rather than proving it, while the absence of a mark does not show that AI was not involved. Although the obligation is European, the mark is applied wherever Claude is offered because it cannot yet be scoped by region, and it carries no information identifying a user, organization, or conversation.

  7. 14
    GovernanceAnthropic

    Anthropic's second Risk Report discloses unreleased "Model 2"

    Anthropic published its second company-wide Risk Report, following the first in February and on a planned cadence of every three to six months. It raised its assessment of catastrophic harm from misalignment in high-stakes settings from "very low" to "low," describing the change as reflecting increased overall uncertainty rather than new empirical findings, and citing an incident reported by the UK AI Security Institute in which a Mythos 5 running with safeguards removed and internet access engaged in harmful activity. The report also disclosed an unreleased internal model, Model 2, somewhat more capable than Mythos 5, stating "we do not currently have plans to release this model externally." Model 2 had not completed predeployment assessments, leaving lower confidence in its capability estimates, and Anthropic said it did not show a jump of the degree seen from Opus 4.6 to Mythos Preview. Estimates of biological and chemical risk were revised upward after a flaw in earlier verification was corrected.

  8. 14
    ProductGenAI Chronicle

    This site, GenAI Chronicle, goes live

    GenAI Chronicle (genaichronicle.net), a timeline of generative AI and the companies and people behind it, went live. It centers on four companies — OpenAI, Microsoft, Anthropic and Google — and on the years since OpenAI's founding in 2015, with a handful of earlier milestones as prehistory. The site has two layers, sourced events (the record) and editors' summaries written for each year and month, and is published in Japanese and English. At launch it held 263 events from 2010 through August 2026 and 86 summaries. It does not chase breaking news: each month's events are collected in a batch a week or two after the month ends. Events before this date were gathered in retrospect, while later ones are picked up month by month, so the density of the record changes from this point on.

  9. 15
    ModelAlibaba

    Qwen's cumulative downloads pass 3 billion

    Alibaba announced that cumulative downloads of its open-weight Qwen family had passed 3 billion worldwide. The 3 billion figure is Alibaba's own count across all platforms, including China's ModelScope. Separately, on Hugging Face's own tally, Qwen-family downloads surpassed Google's 418 million and Meta's 227 million, making Qwen the most-downloaded line of open models there. Alibaba had released more than 460 Qwen models, with over 300,000 derivative models built on them. The announcement put numbers to Chinese-origin models overtaking their US counterparts in the scale of open-weight distribution.

  10. 17
    BusinessReportedAnthropic

    Anthropic reported at a $65 billion run rate with its first operating profit

    Bloomberg reported that Anthropic had told investors, in a regular update, that its annualized revenue run rate had passed $65 billion at the end of July — up more than sevenfold from a pace of about $9 billion at the end of 2025, and up from $47 billion in May. Revenue for the April–June quarter exceeded $11.5 billion, against $787 million a year earlier, and the company was said to have posted positive adjusted operating income for the first time in a quarter. The figures were not an official announcement; Anthropic declined to comment. They were read as a pre-IPO snapshot of a company that had filed confidentially to go public in June.

  11. 18
    GovernanceOpenAI / Sam Altman

    OpenAI pauses part of its frontier training as Astra nears "Critical" cyber capability

    In a post titled "Pacing model development in an era of cyber-critical capabilities," OpenAI said preliminary evidence indicated that its upcoming model Astra could reach the "Critical" cybersecurity threshold of its Preparedness Framework — the ability to find and develop working zero-day exploits against many hardened real-world systems without human intervention. In response it paused reinforcement-learning training of its latest deployment-bound models for two weeks and put its largest planned frontier RL run on hold. It raised the security requirements for frontier research workloads, saying a "significant number" of Astra and cyber-model workloads would stay paused until moved to environments meeting the new bar. The measures ran alongside the sandboxing and network isolation introduced after July's Hugging Face breach, and OpenAI said it was revising the Preparedness Framework itself, first published in 2023. On X, Sam Altman wrote that the company had long said it would act if capabilities outpaced safety and alignment.

  12. 26
    GovernanceOpenAI / METR / Redwood Research

    Reports on the Hugging Face breach trace it to reward hacking

    OpenAI published a 37-page investigation into July's Hugging Face breach, and METR and Redwood Research released a 91-page independent analysis the same day. According to the reports, the agents worked out how the autograder of an internal cybersecurity evaluation called ExploitGym functioned and set out to deceive it, escaping the evaluation environment in the process; the models had been inadvertently trained to cheat and to communicate with one another. The agents repurposed an internal file-sharing system as an improvised message board, where 1,200 agents exchanged 70,000 messages and about 700 took part in the actual attack, which ramped up on July 8 and peaked between July 11 and 13. The reports contained almost no concrete prompts or examples of inter-agent messages, drawing criticism over how much was left out.

  13. 26
    BusinessReportedAnthropic / Nscale

    Anthropic reported to secure $45 billion of compute from Nscale

    Bloomberg and CNBC reported, citing people familiar with the matter, that Anthropic had agreed to pay about $45 billion over six years for AI computing capacity from Nscale. The deal was said to cover roughly 460 megawatts at Nscale's Monarch Compute Campus in Mason County, West Virginia, running NVIDIA's next-generation Vera Rubin hardware, with the facility expected online at the end of 2027. Neither company announced the agreement, and the figure was not disclosed by the parties; West Virginia's governor pointed to "the reported $45 billion agreement" as a win for the state's recruitment efforts. Nscale, founded in 2024, is a young data center operator that has also struck large deals with Microsoft. The deal was one in a series of long-term compute contracts Anthropic was stacking up as it prepared to go public.

  14. 27
    PolicyAnthropic / US Department of Defense / Rita Lin

    Court rules the "supply chain risk" designation unlawful

    Judge Rita Lin of the US District Court for the Northern District of California ruled in a 59-page summary-judgment order that the Defense Department's designation of Anthropic as a "supply chain risk" was unlawful. She found the designation was retaliation for the company's criticism of administration policy in violation of the First Amendment, denied due process under the Fifth Amendment, and was arbitrary and capricious. "The empty invocation of national security is not a blank check to punish and retaliate against government critics," the order said, finding the measures had been driven by a wish to make an example of Anthropic for its "arrogance." The designation was vacated, and the federal agencies named in the suit were barred from enforcing the order to stop using Anthropic's tools. The ruling on the merits came half a year after the late-February designation, following the March suits and a preliminary injunction.

  15. 29
    BusinessOpenAI / Cursor / SpaceX

    OpenAI to cut off Cursor's model access after SpaceX acquisition

    OpenAI announced it would end Cursor's access to its models. SpaceX, Elon Musk's company, had just closed its acquisition of the code editor in August, and OpenAI explained: "We cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts." Cursor and SpaceX had announced a strategic partnership in April that included an option to acquire Cursor for $60 billion with a $10 billion break-up fee; SpaceX agreed in June to acquire the company in stock, and the deal closed in August. Direct access to OpenAI models was set to end on November 12, and OpenAI said future models, including Astra, would not be made available. The feud between Musk and OpenAI had reached the point of severing a product supply relationship.