Friday delivered a rare combination: a cryptographic bombshell from DeepMind, a consumer-friendly video model upgrade, a fintech-flavored Claude feature, a 594B-parameter model heading for open weights, a new agentic workspace from Perplexity, a decoding trick that makes robot brains faster, and the first humanoid robotics IPO on Nasdaq. Here are the seven stories worth your morning coffee, with official sources anchored for every item.
1. DeepMind: AI agents uncover two weaknesses in legacy 1024-bit RSA
Google DeepMind announced on November 6 that AI models, guided by human cryptanalysis experts, uncovered two significant weaknesses in legacy 1024-bit RSA cryptosystems in a matter of weeks. One line of work points to bypassing older digital signature protections; the other involves factoring legacy 1024-bit RSA keys in bulk. Researchers cited in the announcement describe signature forgery becoming roughly 1,600 times more efficient and the marginal cost of factoring additional keys in large batches dropping by about 30 times.
Pushmeet Kohli, DeepMind's VP of Science, was careful to draw the boundary: the results do not affect modern encryption schemes such as RSA 2048. Nobody has named an affected production system, and this is a research announcement rather than a report of a live compromise. But the framing matters more than the numbers. Cryptographic schemes have historically been trusted because their underlying math was assumed hard; an AI-assisted pipeline that shaves centuries of expert effort into weeks undermines that assumption directly.
For defenders, the lesson is concrete rather than abstract. Legacy 1024-bit keys still exist in embedded devices, old certificate chains, and long-lived infrastructure nobody remembers provisioning. AI now gives attackers a cheap discovery engine and gives defenders the same engine, which makes a fresh inventory of legacy keys a sensible weekend project.
— Google DeepMind · The Circuitry
🔗 Google DeepMind research announcements
2. Gemini Omni 1.2 Flash goes GA: 15-second clips, conversational editing, 30% cheaper
Google DeepMind announced general availability of Gemini Omni 1.2 Flash, a speed-and-cost play rather than a capability-max play. Native video generation now stretches to 15 seconds with freely selectable clip lengths between 3 and 15 seconds, pricing drops 30% versus Omni 1.1, and third-party reporting pegs the cost at around $0.07 per second at 720p. The model also handles image, video, and audio references more reliably, with better multi-speaker consistency.
The quietly more important upgrade is conversational editing: a finished draft can be revised by following up with plain instructions instead of rerolling from scratch. For advertising, explainer, and social video workflows, fewer discarded attempts are worth more than extra seconds of footage, because production value lives in whether clothing, characters, and spatial relationships survive multiple revisions.
Distribution is the second half of the story. The model ships through the Gemini API and inside Google Workspace products including Vids, where personal Google accounts get three free 15-second generations per month. That puts reference-driven, editable video one menu away from hundreds of millions of Workspace users, and on the Arena image-to-video leaderboard the model debuted at first place with a score of 1502, ahead of MiniMax H3 and Seedance 2.5.
— Google DeepMind · Arena
🔗 Gemini Omni 1.2 Flash announcement
3. Anthropic launches Personal Finance in Claude, with read-only guardrails
Anthropic rolled out Personal Finance in Claude, a beta feature for US Pro, Max, and Ultra subscribers that connects bank and credit card accounts through Plaid. Users can ask about balances, transactions, and spending patterns in plain language, and the assistant answers from real account data instead of generic advice.
The guardrails are the design. Access is read-only, so Claude cannot move money, pay bills, or manage investments; Anthropic states it does not train on users' financial conversations; and the feature carries an explicit disclaimer that it is not investment advice. That is a deliberately conservative scope for a feature that could easily have shipped with write access and an aggressive upsell loop.
The competitive read is straightforward: ChatGPT launched a similar account-connecting feature in May, so Anthropic is closing a gap rather than opening one. But the strategic stakes are higher than feature parity. Personal financial data carries a different class of trust than chat history, and whoever becomes the interface people ask about money first occupies a position that advertising-supported assistants cannot easily replicate.
— Anthropic · TechCrunch
🔗 TechCrunch on Personal Finance in Claude
4. Thinking Machines reveals Inkling 2 Small: 594B parameters, 17B active, open weights next week
Thinking Machines introduced Inkling 2 Small, a Mixture-of-Experts model with 594 billion total parameters and only 17 billion active, supporting image inputs and variable effort settings. The vision and agentic capabilities are pointed at a specific loop: incorporating visual feedback to iteratively build and refine websites, games, and 3D designs with interactive elements, which is a workload where seeing your own output matters more than raw text scores.
The release strategy is a two-step open-weight play. The model is initially available through Tinker with closed weights while external safety testing and red-teaming complete, and full weights will follow under an Apache 2.0 license next week. That sequence, ship first and open after review, has become the pattern for frontier-adjacent labs that want deployment experience before publication.
The MoE ratio is worth watching for infrastructure teams. With under 3% of parameters active per token, the model is engineered for inference economics, and its evaluation on the Vals Index gives enterprise buyers a third-party reference point beyond lab-published charts. If the open weights land next week as promised, it will be one of the largest Apache 2.0 releases of the quarter.
— Thinking Machines · Vals
🔗 Vals Index evaluation of Inkling 2 Small
5. Perplexity ships Teammates: always-on agents that sit inside your workday
Perplexity launched Teammates, always-on AI agents for work that run continuously alongside a team instead of waiting in a chat window. The positioning puts it in direct competition with the agentic offerings from OpenAI and Anthropic, both of which have spent the quarter wiring agents into subscriptions and enterprise workflows.
The interesting structural question is where an always-on agent lives. OpenAI's approach ties agents into ChatGPT and Codex quotas, Anthropic's into Claude subscriptions and enterprise controls, while Perplexity starts from search and research workflows it already owns. A teammate that has watched your research all week can pick up context that a freshly invoked chat agent has to reconstruct, and that continuity is the product's core bet.
For teams, the evaluation criterion is no longer model quality alone. It is trigger design, connectors, history handling, and failure recovery, which is the same checklist enterprises learned to apply to coding agents this year. The agentic workspace category now has three serious incumbents and a fast follower, which usually means pricing and integration competition rather than raw capability competition.
— Perplexity · The Verge
🔗 The Verge on Perplexity Teammates
6. NVIDIA VFlash makes robot vision fast enough to close the loop, plus four open Cosmos 3 world models
NVIDIA introduced VFlash, a vision model that generates geometric predictions such as robot trajectories and 3D motion in parallel rather than token by token. The numbers: trajectory decoding accelerates up to 12 times, 3D prediction errors drop by as much as 61%, and inference latency falls roughly 80% on Jetson Thor and DGX Spark. On a robot benchmark, downstream manipulation success improved from 62.6% to 70.2%.
The architectural point matters more than any single number. For physical AI, the bottleneck is shifting from model quality to decoding speed, because a control loop can only use predictions that arrive before the world moves again. VFlash uses a lightweight diffusion drafter to emit complete geometric predictions in a single forward pass, a template for making vision-language models fast enough for real-time control rather than just teleoperation demos.
Alongside it, NVIDIA released four open-weight Cosmos 3 models for interactive world simulation and driving scenes, each with 16 billion trainable parameters. They support multi-sensor data generation, synchronized observations, and photorealistic video driven from simulator controls, which turns world models into something robotics teams can download and fine-tune instead of only watching in keynote videos.
— NVIDIA · TLDR AI
7. Agility Robotics rings the Nasdaq bell: the first pure-play humanoid robotics IPO
Agility Robotics began trading on Nasdaq on November 5 under ticker AGLT, becoming the first US-listed, pure-play humanoid robotics company with commercial deployments. The route was a SPAC merger with Churchill Capital Corp XI, completed November 4, at a pre-money equity valuation of approximately $2.5 billion, with about $465 million in gross capital from the trust account, a $201 million PIPE led by Foxconn, and prior equity funding.
What separates this listing from the demonstration-heavy robotics industry is the operating record. Digit robots have accumulated more than 65,000 hours of real-world operation, at a GXO logistics facility in Georgia they have moved more than 100,000 totes, and the company has committed orders exceeding $300 million for the next-generation Digit 5, which runs on an NVIDIA platform and arrives in 2027. Strategic investors include NVIDIA, Amazon, SoftBank, and Foxconn.
CEO Peggy Johnson's playbook is deliberately unglamorous: no backflip videos, focus on repetitive physical work that is hard to staff. Public markets will now grade that discipline quarterly, and the stock becomes the first pure benchmark for whether paid warehouse deployments can support a humanoid balance sheet. A year ago a competitor CEO predicted this company would not survive twelve months; this week it rang the bell.
— Agility Robotics · TechRepublic
🔗 TechRepublic on the Agility Robotics listing
KD Agentic · AI Daily Digest
