OpenAI Sounds the Alarm on Cyber Attacks

On Thursday, Sam Altman declared that the world had reached a “critically important moment for cyber defense with AI” and that there was “not much time to act”. The same day, OpenAI, Anthropic, Microsoft, Google, and Amazon joined more than 100 other organizations in an open letter saying that “in the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated.” Signatories to the letter included major security vendors, banks, and infrastructure operators.

The letter argues that hospitals, water treatment plants, and internet infrastructure are exposed to cybersecurity risk, but that AI now gives defenders a window to fix weaknesses that accumulated over years. It also calls on the government to expedite trusted access programs, which grant selected companies early access to more capable models before general release. It also asks governments to fund cyber defense for under-resourced essential services. Doing so would require a change of course, as the Cybersecurity and Infrastructure Security Agency (CISA) cut a third of its staff last year after budget cuts.

On August 24th, the WSJ published an op-ed by investor Stanley Druckenmiller that was flagged as 100% AI by detector tool Pangram. In response, Druckenmiller said, “Of course I used AI” and “I’m not embarrassed by it… I write everything using AI now for the same reason I use a calculator when I do math problems.” WSJ editorial page editor Paul Gigot defended publishing the piece, noting that outside contributors already lean on speechwriters and that ”what matters is the ultimate argument the author makes and stands behind.”

The WSJ’s defense of AI comes weeks after the FT appended a disclaimer to a column noting “It has come to our attention that AI was used to condense a longer draft of this column prior to submission to the FT and our own editorial involvement” without noting any violation of the publication’s principles around AI, which state “We will not allow AI to compromise the integrity of our journalism, which will continue to be reported, written and created by our journalists and editors.”

While AI authorship may be increasingly tolerated by these publications, other platforms are adopting tools to flag AI writing. Substack now tags user posts with suspected AI content, and LinkedIn shipped a “Seems like AI slop” reader flag, which reduced engagement with suspected AI posts by 40%.

This week, Nvidia agreed to buy Hugging Face, the default distribution point for open-weight models, datasets, and evals, for $12.9 billion. The acquisition values Hugging Face at roughly triple its 2023 valuation of $4.5 billion and ~80x its $150 million annualized revenue. Hugging Face reportedly turned down an investment offer from Nvidia of $500 million valuing the company at $7 billion (though Nvidia did participate in the 2023 round).

It’s the second such acquisition by Nvidia in two weeks. Last week, Poolside sold Model Factory, the system behind the Laguna open-weight coding models, to Nvidia for $6 billion. As part of the sale, over 100 Poolside engineers were offered seats within Nvidia’s open-source Nemotron division. This structure of incumbents licensing tech and absorbing employees while leaving the shell of smaller companies nominally independent is becoming more common. Microsoft’s licensing of Inflection, Google’s licensing of Windsurf, Amazon’s deal with Adept, and Nvidia’s deal with Groq all followed a similar pattern.

• Legora provides a legal AI platform that reached $150 million ARR in Q2 and is in early talks to raise at a $10 billion valuation. Read our new report here.

• OpenAI’s Jalapeño chip beat Nvidia’s GB200 and GB300 systems by 1.5x to 1.9x on throughput-per-watt and by up to 3.6x on latency across three open-weight models tested in-lab by SemiAnalysis, though Nvidia’s HBM4-equipped Rubin platform is the better forward-looking comparison, and Jalapeño remains an engineering sample not shipping in volume until 2027.

• Z.ai claimed the mystery Ox Alpha model that topped OpenRouter’s leaderboard as a new GLM iteration (GLM-5.3-Flash, a 320B-parameter MoE model with MIT-licensed weights released the same evening), and separately said its flagship now runs end-to-end on 100K domestic Chinese chips, sending Hong Kong-listed shares up 8%.

• Anthropic sized its total addressable market at over $30 trillion for IPO investors, edging out SpaceX’s $28.5 trillion May filing figure and roughly equal to US GDP, with the pitch framing the full scope of work AI models could complete as justification for its valuation and infrastructure spend.

• DeepSeek neared a $74 billion pre-money valuation in its second funding round of about 50 billion yuan, targeting an end-August close with CATL, Monolith, Shixiang, CPE, and Legend Capital participating, ahead of a possible Shanghai STAR Market IPO filing by year-end for a 2027 debut.

• Nvidia notified its biggest customers that server prices will rise more than 15% next year on systems containing its flagship Vera Rubin and Grace Blackwell chips, with the size of the increase depending on chip generation and memory configuration as soaring memory costs finally get passed through.

• Tesla recalled ~3 million China-market vehicles over hidden door handles that occupants can struggle to open after a crash, its largest China recall to date and part of a broader 4.3 million-vehicle action across 11 automakers ahead of China’s outright 2027 ban on the electronic-only design.

• New Zealand drafted an Online Safety Bill banning under-16s from Instagram, TikTok, Snapchat, Facebook, YouTube, and X, with fines of up to 10% of global revenue for platforms that fail to verify ages, though the law is unlikely to pass before the November 7 election and would follow Australia’s 2025 model.

• Kalshi released the largest-ever prediction-market calibration study covering 2.2 million resolved markets, finding Brier scores fall from ~0.09 at a three-month horizon to ~0.02 at close, though the study excluded around 80% of markets, including sports, mentions, and intraday markets.

• Amazon hiked hardware prices by as much as 60% overnight across Fire TVs, Echos, Kindles, and Eeros (the $49.99 Echo Dot jumped to $79.99), citing AI-driven memory and storage shortages that Apple’s earlier increases had already signaled would spread across consumer electronics through 2027.

• Oura targeted a September US IPO to raise as much as $3 billion at a valuation above $16 billion, up from $10.9 billion last September, with 2026 revenue projected around $2 billion (roughly 4x 2024’s $500 million) as competition intensifies from Samsung’s Galaxy Ring and Whoop’s reinvented fitness band.

• Meta readied Hatch, its first paid consumer AI agent priced up to $199.99/month, targeting a late-August or early-September launch and running on Claude Opus 4.6 and Sonnet 4.6 until it migrates to Meta’s in-house Muse Spark, with a new Watermelon model planned for October to close the gap with OpenAI and Anthropic.

• China restricted AI companion apps that foster emotional dependence and simulate romantic or family ties, forcing Alibaba and ByteDance to disable major features, with regulators framing the crackdown as a demographic-policy response to a fourth straight year of population decline and record-low marriage rates.

• Huawei pitched Egypt on over 2K Ascend AI chips and a 12-month sovereign-data-center buildout, the first known export of its flagship silicon for military and surveillance use, prompting the State Department to assemble an Nvidia-AMD-Microsoft counteroffer.

• The EPA moved to end the 30-day public-comment requirement for air-pollution permits at “minor source” data centers and the adjacent gas plants that power them, drawing opposition from nearly 200 advocacy groups and more than a dozen states, in the Trump administration’s clearest move yet to blunt the one issue producing bipartisan community pushback against the AI buildout.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!