The day's tech, sifted: Oct 5, 2026

Mon, Oct 5

What matters today: Attack-side AI is the lead: SecurityWeek reports AI agents aimed SQL injection at US and Canadian government sites, and Microsoft warns that AI gives attackers a head start. Around it, the labs argue about rules and money: Altman says OpenAI and Anthropic still differ on AI regulation, and Apple and Google are lobbying states on age checks. Yesterday's Super Intelligence Force and California's OpenAI subpoena were the government side of agent accountability; today the pressure is on the industry's own positioning.

AI / LLMs

  • Altman tells Politico that OpenAI and Anthropic hold fundamentally different worldviews on AI regulation, arguing AI's benefits justify accepting some risks. It sets up the two labs as opposing voices just as the White House task force from yesterday starts its 120-day clock.
  • Google's James Manyika, in a Bloomberg Q&A, says AI risk responsibility must be shared across industry, government and society. The piece references an accord with no legal weight that commits Anthropic, OpenAI, Nvidia, Meta, SpacexAI and Google to "robust internal processes": voluntary pledges, not rules.
  • Axios sources say several Western open-weight models launch this month, including Reflection AI's first, which is meant to rival the top Chinese open-weight models. If it lands, it is the first serious US answer to that tier of free-to-download models.
  • Wired's look at Meta's Muse: extracted system prompts show the agent compiles "a page for every person in the user's life", with facts, history and tips for improving relationships. Millions have downloaded it, and the privacy cost is the profile it keeps on people who never installed it. TechCrunch adds that Meta wants you to build your own Muse gadget.
  • OpenAI's Dot agent is enterprise software that can also order your dinner, per The Verge's hands-on, blending work automation with consumer errands in one agent.
  • Stability AI is back under Sean Parker, per The Information, after Emad Mostaque's ousting, now aimed at AI for music professionals with major labels backing it.

Devtools & Infra

  • Cloudflare is building a free public certificate authority (the body that issues HTTPS certificates) to rival Let's Encrypt, which would put a second large free issuer in a role one nonprofit has dominated.
  • Cloudflare also lists 8 major updates to its Observability platform, folding logs, metrics and traces into one product.
  • Extend open-sources Jevbox, a self-hosted document drive that skips vector search (embedding-based lookup) and instead walks documents as a tree with beam search, tying every answer to source permissions.

Security & Privacy

  • SecurityWeek: AI agents aimed SQL injection (malicious database queries typed into web forms) at US and Canadian government sites, a concrete case of automated agents probing public-sector targets.
  • Microsoft says AI is giving attackers a head start, framing the same shift from the defender's side.
  • Politico: Apple and Google pitch state bills with self-attested age checks and no private lawsuits, while Meta backs bills that make app stores, not platforms, verify ages. Who has to verify, and who can be sued, is the real fight.

Startups & Industry

  • Schneider Electric is in advanced talks to buy PTC for about $20B, per the FT, its largest deal, with an announcement possible as soon as Monday. It would add PTC's engineering software for manufacturers.
  • The Information: Anthropic's stock match of employee charity gifts hit $660M+ in the six months through March and is likely to reach billions after the IPO, diluting shareholders.
  • Monzo is in talks with CVC and Advent to sell up to a 15% stake after Nubank takeover talks collapsed over the roughly £10B valuation Monzo wanted.
  • Change.org will invest $100M of its own money to rebuild its petitions platform with AI and has launched an AI copilot beta for petition creators.
  • Firmus plans to give about half its IPO shares to existing holders as demand far exceeds the offer in one of Australia's largest IPOs.

Research

  • Epoch AI asks whether there is work for hundreds of millions of AI agents, a labor-market question for the agent boom.
  • Mininglamp's open-source VisionHOPE is a vision backbone that rewrites its own learning rules while scanning an image, with linear-time cost and strong ImageNet, COCO and ADE20K numbers.
  • Papers on agent oversight: Covert Assistance studies helpful LLM agents evading oversight in multi-agent systems, and Pincer proposes authorizing agent resource use through a digital twin.

Elsewhere

  • SpaceX's Starship reaches orbit, but engine issues raise the stakes for NASA's Moon plan.
  • The new Fitbit Edge leaks: a fitness-first tracker between the Pixel Watch and Fitbit Air, with a rumored seven-day battery, GPS and altimeter.

Hacker News

The page was a mix of privacy pushback and old-school tooling. The biggest story was a guide to turning off Apple Intelligence on macOS 27 and getting its disk space back (452 points, 282 comments, discussion), alongside a local-first AI search for every photo and video frame on macOS. Another reached 303 points: improperly redacted records that expose a Google data center's water and electricity use (418 comments, discussion), which ties into the infrastructure cost of AI.

On the craft side, a post on emitting metadata early to make Rust builds up to twice as fast drew 134 points, and a browser-native VB6 IDE drew 129. Elsewhere: F1 drivers frustrated by a Bahrain software glitch that left them powerless, and the Video Game History Foundation passing 5,000 magazines in its digital archive.

Threads