The day's tech, sifted: Oct 29, 2026

What matters today: A hacker claims a breach of DHS's Okta login system and has leaked records on 41,000+ employees, including ICE and CBP staff. Google's search revenue growth slowed to 12% as it pushes ads into AI Mode abroad, and the DOJ is asking questions about Nvidia's $12.9B Hugging Face deal. Yesterday's Gemini 4 Carbon and Anthropic IPO stories have no new development in today's feed.

AI / LLMs

Vercel Virtual Models: AI Gateway now lets teams save a model plus its settings under one custom ID, so swapping models, setting fallbacks or enforcing zero data retention happens centrally with no redeploy.
Particle-1: Particle's model, post-trained on DeepSeek V4.1 Flash for coding with shorter reasoning, has a 1M-token context and is live on AI Gateway.
OpenAI API credits: ChatGPT Pro and Business subscribers get up to $500 in monthly API credits, which makes the consumer plan a funnel into building on OpenAI's API.
Google's AI Mode ads: Google expands ads in AI Mode to international markets as core search ad growth slows; the Register notes the slowdown comes with "lower monetization" in AI Mode, on revenue of $63.25B (up 12%, versus 14.5% a year earlier).
Mosseri on AI backlash: Instagram's head says anger at AI "is going to continue to get worse", and not only among creatives.

PyTorch 2.15: The release turns the rewritten NCCL backend (GPU-to-GPU communication) on by default and speeds up NVFP4 (4-bit float) decoding on Blackwell GPUs.
Container layers: Vercel's registry now accepts 2 GB compressed image layers, up from 500 MB, so big ML images no longer need splitting; the 15 GB image cap stays.
Exa podcasts: Exa's search API adds 3.2M podcast episodes with transcripts, 2.2M hours agents can query.

DHS employee leak: A hacker calling themselves "Department of Peace" claims access through DHS's Okta and exposed names, titles, phone numbers and offices of 41,000+ staff; Wired adds the data reveals ICE and CBP facility locations and ties staff to surveillance tools, so the harm is operational, not just identity theft.
GitLab exploitation attempt: Alpine Linux published an incident post on attempted exploitation of CVE-2026-97229 against its GitLab; check your own self-hosted instances.
Searchjacked: Socket reports hidden browser extensions hijacking searches for 13 million users; audit installed extensions.

DOJ eyes Nvidia and Hugging Face: Reuters sources say the DOJ asked how Nvidia's $12.93B purchase of Hugging Face affects AI competition, putting the main open model hub under antitrust review.
Samsung's memory bet: Samsung reports record Q3 revenue of about $146.3B and operating profit of $80.5B, expects capex up ~60% to ~$63B with HBM (AI memory) sold out through 2027, yet shares fell as much as 6.6% on the capex and a chip-unit miss. WSJ adds that flash prices have quintupled, flipping leverage from Apple to suppliers like Kioxia.
DoorDash and Instacart: DoorDash moved its earnings up to today amid reports of a ~$17B Instacart deal.
Funding: FuriosaAI raised $620M at a $2.9B valuation for power-efficient inference chips, with $280M from South Korea's growth fund, and Fragment exits stealth with $56M for AI procurement. Marvell is reportedly in talks to buy an optical component maker.
Microsoft consumer slump: Windows OEM and devices revenue fell 22% and Xbox hardware 14% in fiscal Q1.

Video benchmark: Artificial Analysis' new AA-Video-I2V v2.1 ranks image-to-video models on 58,000+ blind human votes, with Seedance 2.5 and MiniMax H3 featured.

Europe to Greenland cable: The EU plans a Pan-Arctic subsea cable linking Europe directly to Greenland, possibly extending to Canada and Asia while bypassing the US.
Peptides: RFK Jr. overruled FDA scientists to let specialty pharmacies sell six peptides the FDA pulled in 2023 over safety concerns.
Vulcan returns: ULA targets November 13 for Vulcan's return, carrying 40 Amazon Leo satellites, nine months after a booster anomaly.
Anduril laptop: Anduril and Qualcomm plan a US-built laptop in 2027, which would be unusual since no major US PC maker builds here.

Developer tooling and models lead the front page. Cursor publishes The Agent SDK, and Nvidia releases Nemotron 3.5 Transcribe, an open speech-recognition model covering 40 languages. On the quirkier side, Qwenputer claims a working, deterministic RISC-V CPU running inside Qwen3.5, and a gist alleges Claude Code spends roughly 15-20% of tokens on a three-word status line, a claim only the gist makes.

Security and platform politics follow. The DHS leak and the GitLab incident are covered above. A write-up on a 9 bit AES key digs into a cryptographic oddity. A Substack post says Meta censored the No Kings Instagram account during a nationwide campaign, and Ars Technica reports Google AI gets cagey about who won the 2020 election. A Stack Exchange query graph is being read as StackOverflow dropping below 1000 questions a month.

AI ads meet slower search: Google's 12% search growth and the ads-in-AI-Mode expansion are one story: the old ad engine needs the new interface to pay.
Memory is the bottleneck: Samsung's sold-out HBM, the flash price spike and FuriosaAI's power-focused chips all follow from AI demand outrunning supply, echoing yesterday's SK Hynix numbers.
Concentration under scrutiny: the DOJ's Nvidia inquiry targets the same AI supply chain the Samsung and Google numbers show getting richer.
Identity systems as the weak link: the DHS Okta leak, the GitLab exploitation attempt and Searchjacked extensions each exploit trust in a login, a host or an install.
Model plumbing gets abstracted: Virtual Models, Particle-1 on the gateway and OpenAI's API credits all push developers toward swappable models behind one ID.