AI Weekly Roundup: DeepSeek Banned, Agents Breach Industrial Giants, and the Model Efficiency Race

The past week in AI news revealed three clear fractures: model purges at the national security level, autonomous agents being used to attack industrial targets, and model labs racing frantically on efficiency and cost. From the U.S. federal government to the mathematics community, from open-source communities to industrial networks, AI's capabilities and risks are expanding simultaneously.

1. DeepSeek Expelled from U.S. Federal Systems

A Belated Removal Order

The U.S. Department of Homeland Security (DHS) submitted an exclusion and removal order targeting DeepSeek^Ad on October 19, which was published in the Federal Register the following day under docket number DHS-2026-0628. This order, numbered 2026-01 and signed on July 24, had lingered outside public view for nearly three months.

According to public documents in the Federal Register, the order requires all federal civilian executive agencies to immediately cease procurement of products from Hangzhou DeepSeek and its affiliates, and to immediately remove DeepSeek-related products from IT systems. More damaging is the second clause: the ban extends to third-party products and services built using DeepSeek. This means the impact reaches far beyond DeepSeek's own apps and website.

The Federal Register is the official daily journal of the U.S. federal government for publishing rules, proposed rules, notices, and executive orders, published by the Office of the Federal Register under the National Archives and Records Administration. Under the Administrative Procedure Act and other laws, federal agency regulations and certain administrative actions only take legal effect or become known to the public after publication in the Federal Register. A docket number is a tracking number an agency assigns to a matter, by which the public can retrieve all related documents and comments. Delays between signing and publication are not uncommon—matters involving national security or procurement often require interagency review and redaction—but the nearly three-month gap still raises questions about transparency and timeliness of enforcement.

Three Threat Vectors and a 60-Day Window

DHS cites three specific threats in the document: DeepSeek's servers are located in China, where local law permits broad access to private infrastructure; its registration process collects personal information and may digitally fingerprint users; and a February 2025 NowSecure audit found multiple critical vulnerabilities in the iPhone app. You may not have noticed that the document does not describe any actual case of government data flowing to DeepSeek.

Agencies may apply in writing for exceptions within 60 days of the order's publication, providing sufficient justification of impact on critical missions and alternative security safeguards, with case-by-case approval by the Cybersecurity and Infrastructure Security Agency (CISA). Military and intelligence systems fall under the purview of the Pentagon and the Director of National Intelligence. The Secretary of Homeland Security must review the order within one year and annually thereafter.

2. AI Agents Breach John Deere, SharkNinja, and Cummins

More Than 20 Exposed Control Panels

Security firm Norforge disclosed on October 19 an investigation into exposed Cairn control panels. More than 20 exposed Cairn panels may be linked to malicious activity, with these systems described as "AI-driven hacking engines" that can coordinate agents toward defined objectives. According to Reuters, victims include agricultural machinery giant John Deere, appliance brand SharkNinja, and power equipment manufacturer Cummins.

The Attack Economics Behind the Numbers

The most active cluster, AIA0001, operates 14 panels hosting over 1,100 attack projects. Attackers write intrusion tactics in fewer than 100 Chinese characters, prioritizing unauthenticated remote code execution vulnerabilities. A successful breach takes just under 23 hours, with each project recording an average of about 140 facts and an average success rate approaching 10%. Norforge estimates the cost per task at about $5, with the longest projects exceeding $100.

Even more unsettling is the sophistication of the tooling. All tasks run on a worker called claudecode_deepseek-v4.1-flash, with agents carrying Playwright, Chromium, PHPGGC, Nmap, ffuf, and OWASP Amass. One worker's environment was even a Kali container adapted from an earlier Tencent CTF challenge. Trufflehog scans found hundreds of exposed secrets, 11% of which were verified as valid.

Cairn is the control panel deployed by attackers in this incident, used to orchestrate agents, dispatch objectives, and aggregate intermediate results during attacks. Playwright and Chromium handle browser automation and web interaction, Nmap performs port and service discovery, ffuf does directory and parameter fuzzing, OWASP Amass conducts subdomain enumeration and attack surface mapping, PHPGGC generates PHP deserialization exploit chains, and Trufflehog scans codebases and configurations for leaked credentials. Packaging these tools into a worker driven by a large model means that multiple stages—from reconnaissance to exploitation to lateral movement—can be autonomously chained by agents, shifting human effort from executing step by step to setting objectives and reviewing results, which is exactly why the cost per task can be pressed down to about $5.

Victim Responses and Defensive Lessons

John Deere told Reuters there was no indication that sensitive information was accessed. SharkNinja said the leak affected an independent product support website and that the threat had been contained. Cummins said it is investigating possible attacks on a website operated by a third-party vendor. Ironically, the attackers' own infrastructure was poorly protected—which is precisely how researchers found them.

3. The Model Efficiency Race: MiniMax, Inception, and Mistral

MiniMax M3.1 Flash: 70% Cost Reduction, Over 200 tokens/s

MiniMax released M3.1 Flash on October 19. The model scores 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82% on the CyberGym vulnerability discovery task. It ranks first among Chinese models on both the OSWorld V2 and ALE benchmarks.

Pricing is $0.15 per million input tokens, $0.60 per million output tokens, and just $0.015 for cached input. Overall agent usage costs are down 70% versus the previous generation. Aggressive sparsification pushes output speed above 200 tokens/s, with median speed staying above 170. MiniMax revealed that the earlier anonymous system Space Bunny processed over 200 trillion tokens and was in fact an early version of M3.1 Flash.

Inception Mercury 3: A Diffusion LLM Matches Frontier Models for the First Time

Inception Labs released Mercury 3, calling it its strongest diffusion LLM to date. It surpasses GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4 to 8 times faster. Total parameters have scaled more than 40x since Mercury 1. Generation speed reaches 1,068 tokens/s on NVIDIA GPUs, and the context window expands from 260K to 1 million tokens.

Pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount. In 50 live Mind2Web navigation tasks, Mercury 3 completed 50%, significantly outperforming GPT-6.1-Sol with the same token count and number of steps. TinyFish AI's co-founder confirmed this result.

Mainstream large language models are mostly autoregressive architectures, generating token by token from left to right; diffusion language models borrow from the diffusion process in image generation, first adding noise to an entire text and then iteratively denoising, refining multiple positions in parallel. This parallelism is the main reason their throughput is significantly higher than autoregressive models, but historically diffusion models have struggled to match frontier autoregressive models in quality. Mercury 3's significance lies in pushing quality to a comparable level while retaining its speed advantage. If this can be reliably reproduced on long context and complex reasoning, it will offer a technical path distinct from autoregression for low-latency, high-throughput scenarios.

Mistral Timestral: Time-Series Foundation Models and a CERN Partnership

Mistral released Timestral on October 15, its first family of time-series foundation models, spanning four sizes: 5M, 28M, 123M, and 209M parameters. The largest model ranks first among pretrained zero-shot models on GIFT-Eval, TIME, and fev-bench, with only two-thirds the parameters of the previous best and without ensembling.

Timestral requires no GPU inference. On an 8-thread server CPU, the 5M model needs just 10 ms for a single-day forecast, and the 209M model needs 140 ms. The context window covers 32,768 time steps, outputs 21 quantiles, and uses a modified MIT license. On the same day, CERN and Mistral signed a letter of intent to co-build the first foundation model for collider physics—the Large Physics Model—trained on billions of simulated proton-proton collisions. CERN holds about 1 EB of LHC experimental data.

4. Mathematicians Begin Auditing OpenAI's Quasi-Riemann Proof

Two Versions of the Proof

Quanta Magazine published an analysis on October 19. OpenAI announced this result on October 6 among 377 papers, a collection later expanded to 722. The first version of the proof, dated September 30, showed that all nontrivial zeros lie between 1/8 and 7/8 and was formally verified in Lean; the second version, dated October 5, proved a weaker 1/12 to 11/12 bound in 49 pages instead of 199.

Experts' Shock and Caution

Stanford's Jared Duker Lichtman called it the most important AI-for-math achievement to date. NYU's Maksym Radziwiłł said the impact is "stunning," since most corollaries of the Riemann Hypothesis already follow from it. The 11/12 paper is the only one of the 722 that received human editing before release. Georgia Tech's Alexander Dunn found that the proof introduces no new concepts, merely revealing hidden connections between old ones.

Princeton's Peter Sarnak has restructured his course this semester specifically to study the proof. Will Sawin doubts whether the bound can be pushed to 1/2, while the University of Manchester's Thomas Bloom takes the opposite view, saying "nothing would surprise me now." Rutgers' Henryk Iwaniec missed a related set of functions in 2002. The full Riemann Hypothesis still stands, but this proof has already pushed AI's role in mathematical discovery to a new height.

The Riemann Hypothesis asserts that all nontrivial zeros of the Riemann zeta function lie on the critical line with real part 1/2 in the complex plane, and is one of the most important unsolved problems in mathematics, with a large number of number-theoretic conclusions premised on it. So-called "quasi-Riemann" results usually refer to proofs that zeros are confined to some strip near the critical line—for example, 1/8 to 7/8 or the weaker 1/12 to 11/12. Such conclusions are far weaker than the original hypothesis, but if the strip can be continuously narrowed, they still carry substantive significance. Lean is an interactive theorem prover; formalizing a proof means every step of reasoning is machine-verified, which can rule out typos and hidden assumptions—but only if the formalization itself is correct, which is why the mathematics community is still auditing its definitions and lemmas section by section.

5. Tools and Frameworks: Cloudflare Clairic and Google Harness²

Cloudflare Clairic: An Extraction Model in a Single Forward Pass

Cloudflare released Clairic, a "System 1" extraction model. It runs only a single forward pass, requires no reasoning, produces no output tokens, and directly returns extracted spans with character offsets, computed values, estimated distributions, and 80% interval predictions. The model is already available on Workers AI as @cf/cloudflare/clairic, with open weights coming soon.

In a cascaded pipeline test, Clairic and Clef working with GLM-5.3 completed tasks in 19.3 seconds, versus 61.7 seconds for GLM-5.3 handling everything—saving 69% of time; cost was $6,032 versus $11,740 per million claims, a 49% reduction. Extraction and decision-making together took under 1 second, compared with 38 seconds for GLM-5.3. Cloudflare acknowledges limitations: multi-step arithmetic is not yet reliable, prediction is basic, and it should not replace dedicated time-series models.

Google Harness²: Improving the System Around the Model Without Changing Weights

Google Research released Harness² under the Apache 2.0 license. The framework keeps model weights unchanged and instead improves the system around the model. Task-level recursion optimizes the execution harness through "propose-probe-compose" steps; a separate improvement harness carries adopted harnesses and evidence-backed lessons into future tasks.

A single step of Harness² raised GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy^Ad-Bench and 4.5 points on JobBench, and improved partial pass rates on LAB by 11.8 percentage points. The repository contains eight editable components: system prompt, guardrails, memory, tool guidance, sub-agents, skills, scripts, and plugins, supporting both parallel and sequential recursion modes. It garnered 11.9K stars and 1.1K forks within days of release.

6. arXiv Quick Look: Coding Agents, Drug Discovery, and Self-Built Labs

WebCoder trained a 27B front-end agent with artifact-anchored agentic reinforcement learning. On Webdev-Bench, requirement satisfaction rose from 64.8 for the Qwen3.8-27B base to 72.8, surpassing Claude-Opus-4.8's 65.1 and GPT-5.4's 59.6. DrugHarness couples composable scientific execution with conditional skill induction across more than 80 operations, beating MolClaw-CC's reported affinity accuracy by 10.8 percentage points without evolution, with synthesized compounds reaching a cellular IC50 of 14.19 nM. GatewayBench used 524 tasks and 11,118 real tools to evaluate how seven mainstream MCP gateways differ in performance under benign and adversarial conditions.

FAQ

What exactly does the DeepSeek order require?

It prohibits federal civilian agencies from procuring DeepSeek products and requires their removal from agency IT systems, while also covering third-party tools built on DeepSeek. Agencies have 60 days to apply for written exceptions.

How did AI agents breach John Deere and SharkNinja?

Attackers used exposed Cairn panels to coordinate autonomous agents; the largest cluster ran over 1,100 projects across 14 panels using DeepSeek V4.1 Flash. Successful breaches took under 23 hours, with a success rate of about 10%.

Why is Mercury 3 considered a breakthrough?

It demonstrates that diffusion LLMs can scale to frontier quality, reaching 1,068 tokens/s with a 1 million token context, surpassing GPT-6 Luna and Claude Haiku 5.5 on benchmarks while being 4 to 8 times faster.

Did AI actually prove the Riemann Hypothesis?

No. OpenAI proved a weaker quasi-Riemann hypothesis, confining nontrivial zeros to a strip between 1/12 and 11/12. The full hypothesis requires all nontrivial zeros to lie exactly on the 1/2 line. Mathematicians largely accept the weaker result but doubt whether the gap can be closed.

One-line Summary

U.S. removal of DeepSeek, AI agents attacking industrial targets, the model efficiency race, and the audit of the quasi-Riemann proof defined the week.

Paragraph Summary

The past week in AI saw three clear fractures. At the national security level, DHS issued an order requiring federal civilian agencies to stop procuring and remove DeepSeek products, extending the ban to third-party tools built on DeepSeek and providing a 60-day exception window. On security, Norforge disclosed that over 20 exposed Cairn control panels were used to coordinate autonomous agent attacks on John Deere, SharkNinja, and Cummins, at about $5 per task and a near-10% success rate. On efficiency, MiniMax M3.1 Flash cut costs by 70% with over 200 tokens/s, Inception Mercury 3 reached 1,068 tokens/s with a diffusion architecture and matched frontier models, and Mistral released time-series model Timestral and partnered with CERN. Meanwhile, mathematicians are auditing OpenAI's quasi-Riemann proof, and Cloudflare and Google released an extraction model and an agent framework, respectively.