LONG AI
Home › Blog › DeepMind Agent Finds Two Weaknesses in 1

DeepMind Agent Finds Two Weaknesses in 1024-Bit RSA

Category:行业资讯 · 17 min read · 2026-11-08

This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Headline: DeepMind agent finds two weaknesses in 1024-bit RSA

DeepMind reported that an AI agent found two weaknesses in 1024-bit RSA, and - the sharper number - reproducing the same attack dropped from 1,380 core-years to 391 core-years, completing in just 3.2 days.

The significance isn't "AI cracked crypto" (1024-bit RSA is already considered marginal for high-security use), but the efficiency gain: the agent compressed the search enough to make a formerly prohibitive computation feasible. This is AI doing work on problems with no standard answer, not leaderboard score-gaming.

Other releases (6)

Model / ProductDateTypeHighlights
Arete-27B (Nous Research, open source)11-6Value post-training / full weightsPure SFT on Alibaba Qwen3.8-27B - no architecture changes, no RL, no preference optimization, only ~52.8M tokens. Goal is changing "stance," not "capability": on a self-built 803-item value eval, official-wording rate dropped from 22% to 0.9%, fact score 2.32 to 2.98. Capability nearly untouched (Hermes Bench 65.60 vs 65.77). Teacher was Thinking Machines' Inkling (open model to dodge closed-source distillation terms). Apache 2.0, weights + eval set + training data all released; BF16 ~53.81GB, Q4_K_M 17.20GB runs on 24GB. Honest caveats: 216 of 803 still score 0; self-eval, no third-party rerun; text layer only. Technical value over product value - it proves "changing values" now takes tens of millions of tokens, and the recipe is public.
Qwen Concept 27B (varien / Concept Country, open source)11-6Weight-space editing (no training)Claims the first "no training, only edit weights" open model: a general model reads and rewrites another network's weights without gradients or training loops; BF16/FP8/NVFP4. Vendor: -54% thinking tokens, GPQA Diamond +3%, +30% tokens/joule. Background is MIT CSAIL's Nov 2 paper (unreviewed) using GPT-6 Astra to edit just 0.016% of Qwen3.5-0.8B weights, lifting MATH-500 7.9% to 48.1%. Three flags: all self-reported; weight links unverified at announcement; no license - "open source" is still "announcement stage."
Drex decision model (Nace AI, open source)11-6Decision-specialistSame route as Kakao and Ant: no text generation, only per-option probabilities. The 27B sits 1.6 points from Claude Opus 5.5 on Decision Bench v4, 183ms response (vs 3.5s); RLAF adds 2.0 points. For routing, moderation, scoring - "a judgment, not a paragraph."
Anchor 1 (Archon Robotics)11-7Full-body control / embodied "cerebellum"A 32-page tech report, positioned as System 0 - the layer translating commands into full-body motion. The trick: it only outputs five anchor points (pelvis, wrists, ankles position/orientation), leaving the rest to controllers; human mocap is translated offline, not per-frame online retargeting. A pragmatic engineering route in embodied AI.
Hugging Face open MoE training guide11-7Engineering guide / open code~800K GPU-hours of experiments distilled into ~70K words and 55 interactive charts: expert allocation, communication latency, load balancing, routing numerical stability (route in FP32), forced-uniform vs learned routing. Systematizes the community's scattered know-how for MoE fine-tuning.
Meta Muse proactive memory to consumers11-7Product / memoryBrings the enterprise memory-specialist model to Muse and Meta AI: proactive memory and proactive focus. Vendor: correct key-context use 76% to 90%, key-info forgetting down ~2.4x, recovery after constraint changes +67%, internal "personal superintelligence index" 38% to 47%. Plus Spaces (describe a persistent mini-app). Vendor-reported.

Industry news

The trend in one line

AI is now compressing "impossible-scale" computations into "practical" ones - the frontier is efficiency, not just capability.

(Data from official announcements and public reports, 2026-11-08; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa