Four developments today that span the full stack: cryptographic infrastructure, frontier model competition, productised video generation, and the first US humanoid robot stock. Each carries a decision for operators planning the next twelve months.

DeepMind AI agents find two weaknesses in legacy RSA-1024

On 6 November, Google DeepMind's Pushmeet Kohli announced that AI agents, working alongside human cryptanalysts, had identified two algorithmic weaknesses in obsolete 1024-bit RSA schemes. The first concerns digital signature forgery: when an attacker controls approximately half of the signed data block, the RSA padding scheme fails to provide adequate protection. The second is a batch-factoring acceleration, where a new algorithm with better asymptotic complexity cuts the average cost of factoring each additional RSA-1024 key (once one has been solved) by roughly 30 times compared to previous batch methods.

The scope is limited. Kohli stated that "modern encryption protocols, which almost universally use 2048-bit or larger keys, remain secure." But RSA-1024 persists in payment gateways, embedded certificates, legacy PKI chains, and older IoT firmware. If your organisation has not yet audited key lengths across its full vendor and infrastructure stack, now is the time. The cost of migration rises as these attack paths become cheaper and more accessible.

Ant Group releases Ling-3.2-flash, a 560-billion-parameter multimodal model

Ant Group launched Ling-3.2-flash on 6 November: a 560-billion-parameter mixture-of-experts model that activates 25 billion parameters per token and supports a one-million-token context window. The model handles text, images, and video natively in a single forward pass, without a separate vision encoder stage.

Ant's own benchmarks claim it outperforms Claude Opus 5 on three visual evaluation sets: OmniDocBench 1.5 (92.21 vs 90.56), ZEROBench-Sub (43.71 vs 39.52), and Vision2Web (85.86 vs 80.20). It is free to use for the first two weeks, then priced at $0.30 per million input tokens and $0.90 per million output tokens. That puts Ling-3.2-flash at roughly one third the output cost of Claude Haiku 5.5.

Treat the benchmarks with the standard caution: Ant selected the evaluations, and no independent text-only comparisons have been published. The right response is a targeted trial on your document-extraction and UI-agent workflows before drawing conclusions. The pricing point alone makes the experiment low-cost.

Gemini Omni 1.2 Flash goes generally available with longer clips and lower prices

Google made Gemini Omni 1.2 Flash generally available on 6 November across the Gemini API, Google Vids, and Google Cloud. The main changes from 1.1: clips extend to 15 seconds (from 10), 720p video drops 30 per cent to $0.07 per second, and conversational editing now lets teams refine the same clip without resetting the generation. Supported refinements include swapping a product, changing lighting, or adjusting camera angle. Outputs carry a SynthID watermark. Provisioned Throughput is available for enterprise volume commitments.

Requests to the Omni 1.1 Flash endpoint will begin routing to 1.2 from 3 December. For teams already using the API, the main action is to verify that 15-second clips do not break downstream assumptions about clip duration, and to review billing at the new rate.

Agility Robotics lists on Nasdaq as the first US humanoid stock

Agility Robotics (ticker: AGLT) began trading on Nasdaq on 5 November through a SPAC merger with Churchill Capital Corp XI, closing its first day at $11.51. The merger values the company at approximately $2.5 billion. Backers include Nvidia, Amazon, and SoftBank Vision Fund 2. The company reported $1.8 million in revenue last year against a net loss of $138.1 million, and has secured more than $300 million in multi-year orders for its Digit 5 robot. Deliveries are targeted for the first half of 2027, with general availability by the end of that year.

Agility is not the global first (Chinese firm Unitree listed in Shanghai in August), but it is the first in the US. The fundamentals are unambiguous: this is a pre-revenue-at-scale business going public on the promise of a specific deployment timeline. The signal is not the share price. It is that institutional capital is now underwriting a 2027 delivery cycle for warehouse and logistics automation. Operators planning in that space should track the order-book fill rate and pilot-access announcements, not the stock movement.

The through-line across today's brief is timeline compression. AI is accelerating attacks on legacy security infrastructure, closing benchmark gaps between Chinese and Western frontier models, maturing video APIs from preview to production pricing, and moving physical automation past the public-market threshold. Each of these has a decision attached; none of them can be watched from a distance.