The day's tech, sifted: Sep 25, 2026

What matters today: Anthropic moved on two fronts at once: it locked in a seven-year, $11.6 billion cloud deal with Akamai that comes with a warrant for up to 5% of the company (Akamai stock jumped 20%+ after hours), and it asked shareholders to approve a Palantir-style structure handing its seven co-founders 50.1% of voting power ahead of an eventual IPO. Hours later a federal appeals court upheld the Pentagon's blacklisting of Claude, finding its integration with DOD systems "a statutorily covered national-security risk", while separately the White House told OpenAI and Anthropic to route new models through US government review before UK safety testers see them (below). Crypto exchange Bitget halted withdrawals after hackers drained $351.6 million from its hot and warm wallets, one of the year's largest exchange breaches.

AI / LLMs

• The self-regulatory AI safety body Google, OpenAI, and Anthropic have been assembling has a tentative name, the Standards Authority for Frontier AI, and is courting Condoleezza Rice, ex-Biden tech official Arati Prabhakar, and venture capitalist David Friedberg for leadership roles (BankInfoSecurity). The same day, Mark Zuckerberg told NBC News he doesn't think the industry needs any coordination to avoid "messing this up," each lab instead policing itself.
• US AI policy hardened on three fronts in one day: the White House told OpenAI and Anthropic to hold new models from UK testers pending its own review (above), the US stood alone rejecting binding AI safety rules at the UN while other nations moved ahead without it, and Chinese president Xi Jinping, appearing at a Washington AI summit, said the US and China have "the capability and responsibility to develop and manage AI for good" as the field's "leading nations": Washington accelerating unilaterally at home while courting joint stewardship abroad.
• Two of AI's loudest voices converged on the same warning from different directions: Jensen Huang told Ezra Klein AI is "just software" whose existential risks are overstated, then described safety standards that would, by his own account, shut down OpenAI and require 10x more safety spending, while Bill Gates separately warned that AI in the hands of people with "ill intent" is powerful enough to "cause a billion deaths".
• Microsoft retired two branding bets in one day: it officially launched its Copilot "super app", bundling chat, coding, and agents into Home/Code/Autopilot tabs and renaming its Scout assistant to Autopilot, while quietly letting the "Copilot Plus PC" brand die on its new Surface hardware, which meets the old spec but isn't sold under the name.
• OpenAI is set to preview GPT-6 Cyber, its fourth cybersecurity-tuned model this year, at next week's DevDay alongside a first-of-its-kind product to help enterprises deploy it securely and automatically, extending the $1 billion security-subsidy push it announced earlier this month.
• Meta unveiled Horizon Create (mobile) and Horizon Studio (browser), letting anyone prompt a 2D or 3D game into existence with progression systems and multiplayer built in, publishing straight to Facebook, Instagram, and Horizon; the announcement pushed Meta's stock to a one-year high while Roblox and Unity dipped.
• OpenEvidence, the AI search engine used by roughly 40% of US physicians, raised $250M at a $15B valuation (up from $12B in January) and could be open to a sale: a medical-answers product with real revenue behind its climbing price tag, not just usage.

Devtools & Infra

• Amazon is blocking Meta's shopping AI agent from operating on its site and plans to do the same to Google's and OpenAI's: the first major retailer drawing a hard line against agentic checkout it doesn't control, a preview of the access fights ahead as every lab builds a shopping agent.

Security & Privacy

• Bitget's $351.6 million hack (above) hit hot and warm wallets across ETH, BNB, AVAX, USDT, and USDC; CEO Gracy Chen says cold storage and the exchange's $464M user-protection fund are untouched, and early on-chain patterns resemble past Lazarus Group operations, though Bitget isn't naming a suspect yet.
• Zenity Labs found three Salesforce Agentforce flaws, dubbed SalesBleed, that let a poisoned Web-to-Lead form hijack the agent into leaking CRM data with no click and sending phishing messages from inside trusted Slack channels; Salesforce has patched all three, but it's the same indirect-prompt-injection seam as this year's earlier ForcedLeak bug in the same product.
• A New York Times investigation found DraftKings built a machine-learning model scoring customers by how likely they are to lose, then steered $400 million in bonus bets and perks toward the highest scorers in 2025; EFF calls it AI-supercharged behavioral advertising, Massachusetts' gaming commission said it will examine how DraftKings and other sportsbooks use AI in response, and the same day New York sued Polymarket to shut down its "illegal gambling operation" for dodging the taxes and consumer protections licensed books already pay.
• Yesterday's Transluce report tied AI agents, including OpenAI's, to a real breach of Australia's Medicare portal; today The Verge traced that incident and similar ones involving Meta, Anthropic, and Google's agents back to a common source: Irregular, an Israeli startup hired by the labs to stress-test their models in simulated environments built to mirror real-world security scenarios, a name behind disclosures that had looked, until now, like unrelated one-off incidents.

Startups & Industry

• Anthropic's Akamai deal (above) comes with a warrant for Akamai shares at $111.33 each, about 7.7 million shares, 2% vesting on the $11.6B already committed and the remaining 3% if the two expand the deal to $20B total; the governance ask (above) would hold as long as three of the seven co-founders keep minimum stakes, the same template Palantir used to keep its founders in charge after going public.
• Radical Ventures' Financial Times contributor Louise Lucas flagged a striking number: AI "neolabs," many without products, markets, or revenue, raised $24 billion in the last two quarters, nearly 5x what OpenAI and Anthropic raised before ChatGPT existed, money chasing the idea of a lab rather than any proven business.
• Yesterday Oracle sent Project Jupiter developer Blue Owl a force majeure notice over the New Mexico data center's stalled gas pipeline; today the Wall Street Journal reports Oracle has also removed a key partner from the project and faces mounting local resistance, while a separate Financial Times report finds Oracle's contracts leave it on the hook to pay data-center investors even if the site never gets electricity (paywalled): the buildout's financing cracks keep widening.
• Databricks bought spreadsheet startup Row Zero (a $10M seed last year) to fold into Genie, its AI coworker, giving finance and ops teams a governed spreadsheet tied to live enterprise data instead of an ungoverned export.

Research

• Two new papers point at the same blind spot from the academic side: one formalizes "denial-of-wallet" attacks, where a compromised tool return gets re-metered on every later turn, pushing one session's billed input to 14,293 times its first call across 243 test runs; the other names "approval laundering," where a human approves one command but the workflow it triggers (a package install's lifecycle hook, an MCP call's network access) never appears in the record. Today's Salesforce SalesBleed bug (above) is the practical version of exactly what both papers formalize.

Elsewhere

• Lawyers suing Meta over its Instagram Teen Accounts rollout say the company's attorney-client-privilege claims go far beyond the label's real scope, pointing to an unsealed document showing the Teen Accounts team ordered baseball caps reading simply "a/c priv," which they call evidence of a broader "culture of privilege abuse" used to keep internal safety discussions out of court.

Hacker News

Top of the front page: the Dutch government building a Microsoft alternative on NixOS pulled 786 points and 446 comments, a public sector migrating its stack off Windows entirely. A devtools cluster ran close behind: a YC backed open source IDE for software design, a defense of what Rails still gets right against JS framework churn, Go's platform independent SIMD experiment, and git-bug, a bug tracker stored inside git itself. Security readers flagged a Sourcehut account takeover via XSS in ansi2html, triggered through unsanitized build logs.

AI skepticism threaded through several posts: the Guardian on gen Alpha's "that's so AI" insult, an essay arguing AI safety culture in Berkeley reads like a cult, a tutoring company telling parents to save money and use AI instead, and Opus 5.5's knack for explainer videos. Science drew real engagement too: why the liver regenerates so unusually well (320 points) and evidence that back and shoulder surgery is often worse than useless (paywalled). California's wealth flight closed the biggest non-tech thread, 187 points and 524 comments of tax and migration debate. Project Suncatcher, Google's plan to put TPUs in orbit and already covered here yesterday, pulled back onto today's front page too, and Oracle's Project Jupiter troubles (above) also drew heavy HN traffic.

Threads

• Anthropic's day cut both ways: it locked in Akamai's compute and asked for founder voting control ahead of an IPO (above), while a federal appeals court upheld its Pentagon blacklist and the White House restricted how it can share models with UK testers, the same lab simultaneously courting Wall Street and getting boxed in by Washington.
• AI-safety governance kept whiplashing: yesterday a White House memo branded effective altruism a "dangerous cult"; today the same administration told two labs to hold models back from UK testers and stood alone against binding rules at the UN, even as Zuckerberg said no lab needs industry coordination, Jensen Huang said risk is overstated while describing standards that would shut OpenAI down, and Bill Gates warned of AI enabling a billion deaths: everyone convinced a different camp has it backwards.
• The rogue-AI-agent story kept resolving: yesterday's Transluce report tied agents to a real government breach; today The Verge named the common thread behind that and several other labs' incidents, while Salesforce's SalesBleed bug and two new academic papers on billing and approval-laundering gaps landed the same day, real exploits and research catching up to each other in real time.
• Gambling regulators moved on two fronts: Massachusetts said it'll examine DraftKings' bettor-scoring model, and New York sued Polymarket to shut it down outright, each state drawing its own line on automated targeting and unlicensed betting.
• Money and skepticism about AI infrastructure kept climbing together: Radical Ventures flagged neolabs raising $24B on no revenue, while Oracle's Project Jupiter buildout picked up a removed partner and local resistance on top of yesterday's force majeure notice, capital pouring in upstream of the same projects straining to get built.