Today’s through-line spans cryptanalytic research, multimodal generation, a dense-activation MoE on an open-weight path, scientific-agent evaluation and consumer read-only finance: Google DeepMind reported human-guided AI agents finding two major weaknesses in legacy 1024-bit RSA while stressing that modern 2048-bit and larger keys remain secure; Gemini Omni 1.2 Flash became generally available; Thinking Machines Lab released Inkling 2 Small with pretraining and joint RL detail; OpenAI and Ligo launched BioBasicsBench for routine wet-lab tasks; and Anthropic began rolling out Claude Personal Finance for U.S. paid tiers via Plaid, stressing read-only access and no training use.

1. Google DeepMind: human-guided AI agents expose two weaknesses in legacy 1024-bit RSA

On November 6, 2026, Google DeepMind—framed by Pushmeet Kohli, VP of Science—announced that human-guided AI agents had uncovered two major weaknesses in legacy 1024-bit RSA. The work was led by Dj Dvijotham and Abhradeep Thakurta, in collaboration with Nadia Heninger (UC San Diego and Inria) and Emmanuel Thomé (Inria). DeepMind and related reports emphasize that modern RSA at 2048 bits and above remains secure, and that the signature experiments were sandboxed on an already-solved 1024-bit benchmark without touching live systems.

On signatures and padding, DeepMind stated that once an attacker controls roughly half or more of the signed block, padding adds virtually no protection. In the sandbox, the improved approach used about 3.5× less CPU time and less than one-thousandth of the compute needed to break the key itself. When the attacker controls only a short fingerprint (for example 32 bytes), the new algorithm is about 1600× better than the previous best for 1024-bit RSA on theoretical estimates. Separately, for batch factoring of same-bitlength moduli, DeepMind reported roughly 30× lower average cost per additional 1024-bit key versus prior batch methods.

Takeaway: AI-assisted cryptanalysis is sharpening attacks on legacy 1024-bit RSA under stated assumptions, while DeepMind’s own framing keeps the operational boundary clear: 2048-bit and larger deployments are not claimed to be broken, and the signature tests stayed off live production.

2. Google: Gemini Omni 1.2 Flash generally available — native video to 15s, price down ~30%

Google DeepMind product manager Anish Nangia announced on November 5, 2026 that Gemini Omni 1.2 Flash is generally available (GA) on the Gemini API and Gemini for business / Google Cloud. The model card carries the same publish date. Versus Omni 1.1’s 10-second ceiling, 1.2 Flash extends native video to 15 seconds, with duration selectable from 3–15 seconds in one-second steps; the 720p price falls about 30% versus Omni 1.1, to $0.07 per second.

On capability, the post emphasizes multi-turn conversational editing and stronger adherence to reference images, motion clips up to about three seconds, and audio. Workflows can draft at 360p / 720p then upscale to 1080p / 4K, with SynthID watermarking. Provisioned Throughput is available. The model is also rolling out in Flow, Google Vids, the Gemini app (Ultra), Google Ads, AI Studio, Vertex AI, YouTube and Workspace.

Takeaway: Video generation is shifting from “can it produce a clip” to priced duration tiers, reference adherence and enterprise throughput guarantees, with draft resolution plus upscaling written into the default workflow.

3. Thinking Machines Lab ships Inkling 2 Small: 594B / 17B active, Apache 2.0 weights planned

Thinking Machines Lab announced Inkling 2 Small on November 5, 2026: a mixture-of-experts (MoE) model with about 594 billion total parameters and 17 billion active, supporting image inputs and variable effort. The company cited an Artificial Analysis Intelligence Index score of 31 (the post says 31; a table shows AA Index v5.0 at 30.6). The model launches initially on Tinker with closed weights; the announcement says full weights are planned under Apache 2.0 next week after external safety testing and red-teaming.

Training disclosures include: H1 pretraining of about 8.4 trillion tokens in roughly 68 hours on NVIDIA GB300 NVL72 with native NVFP4; H2 of about 3.8 trillion tokens at up to 512K context; then joint RL for 1,075 steps and more than 17 million rollouts over about nine days on 2,048 GPUs. The lab showed agentic coding, vision and knowledge-work demos, and cited a DesignArena Website score of 1340 in its snapshot.

Takeaway: The “small activation, large total” MoE path binds pretraining throughput, joint-RL scale and a “ship closed, open weights next” cadence into one release narrative.

4. OpenAI launches BioBasicsBench: measuring how AI agents handle routine wet-lab science

OpenAI published on November 6, 2026 an introduction to BioBasicsBench, built with enzyme-design company Ligo. The benchmark covers 100 routine wet-lab tasks across five areas—protein methods, imaging / cytometry, quantification and assays, nucleic-acid methods, and molecular cloning—with inputs such as raw gels, plates, microscopy and instrument exports. Evaluation spanned seven models, 37 settings and five attempts per task, yielding about 18,500 results.

At xhigh reasoning, GPT-6 Astra posted a 66.7% mean reward (95% CI 60.3%–73.1%), GPT-6.1 Sol 62.0%, and GPT-5.6 Sol 43.8%. The company stresses that higher scores do not yet equal reliable lab judgment. Examples include Astra catching a wrong antibiotic on a control-plate rim label, and remaining failure modes such as assumptions overriding controls and, in one case, colony miscounts of about 12% overcount.

Takeaway: Vendors are turning “reading raw experimental artifacts” into a reproducible wet-lab benchmark, while using confidence intervals and failure cases to blunt “high score equals trustworthy” extrapolation.

5. Anthropic: Claude Personal Finance beta rolls out — read-only account links via Plaid

Anthropic began rolling out a Personal Finance beta this week for U.S. Pro, Max and Ultra users; Plaid confirmed the partnership in a November 5, 2026 post, with TechCrunch and others covering the launch. Users can connect banks and cards through Plaid. The capability is read-only: it cannot move, manage or invest money, and carries no ads or sponsored products. Connected accounts can surface up to about two years of transactions; chats in the Personal Finance hub are not used for training; users authenticate at their bank via Plaid so Claude does not see bank passwords; unlinking is available anytime. Anthropic states the feature is not investment advice.

Takeaway: Consumer assistants are entering account-level financial insight, but with read-only, non-training, non-advisory and unlink-anytime boundaries—routing connectivity through payment infrastructure rather than building bank links in-house.

---

One-line takeaway: DeepMind reports AI-assisted advances against legacy RSA-1024 while affirming modern 2048-bit+ security, Google makes Gemini Omni 1.2 Flash generally available with longer native video and a lower unit price, Thinking Machines lays an “Apache 2.0 next week” MoE path with Inkling 2 Small, OpenAI quantifies wet-lab agent performance with BioBasicsBench while warning that high scores are not reliable judgment, and Anthropic rolls out read-only Personal Finance for U.S. paid users via Plaid.