Gemini Enters Enterprise Agency | Rewire Morning News

Agents turn models into swappable engines, yet platforms, compute, verification, and power remain in separate hands.

1 | Google launches Gemini enterprise agents, can call external models like Claude

Google has turned Gemini into an enterprise agent that can take on tasks in work software like Gmail, Drive, and Docs, and allows the underlying system to call its own models or Anthropic's Claude, with intelligent routing and project spending caps. The company says nearly 90% of the Fortune 100 already use Gemini Enterprise, according to Google's own account.

The change is that enterprises can switch models by task, but the agent's context, employee permissions, and workflows remain at the same entry point. What Google is competing for is not just a single Q&A, but the control plane from task assignment to delivery. Multi-model choice reduces the bet on a single model, but may hand a new layer of dependency to the agent platform. What enterprises ultimately buy is not just model call volume, but also the ability to connect organizational knowledge and employee permissions to the agent.

(Source: Google Cloud / Reuters / The Verge)

2 | OpenAI withdraws 3 math preprints, researchers question bulk release process

On October 8, TechCrunch reported that the release of these results deviated from the norms proposed by the group of math researchers OpenAI consulted. Retraction Watch subsequently followed up on the retraction and revision progress. On October 6, OpenAI publicly released 722 math manuscripts generated by internal frontier models, covering 372 sets of results, many with Lean formal proofs attached.

On October 7, the company admitted that a notation error would break through 3 related papers, withdrew those 3, and revised another 14. Repository records show that after revision, about 42% of the 719 top-level results were formalized. Bulk output shifts scarce resources from writing proofs to verification, understanding, and attribution, and also turns error correction speed into a publicly visible part of research.

The focus of the controversy is not only errors in individual papers, but also who invests peer review in hundreds of model outputs. Lean only verifies the formalized parts that machines can check, and does not automatically judge whether results are novel or important, nor does it guarantee that researchers can explain the proof strategy.

(Source: TechCrunch / OpenAI math repository / Retraction Watch)

3 | Anthropic launches critical infrastructure protection plan, and offers free scanning of open-source code

Anthropic expands its AI defense program to industrial control systems in power grids, water utilities, and transportation, partnering with service providers including Accenture, CrowdStrike, and Rockwell to give critical infrastructure defenders access to frontier models, on-site engineers, and threat research. It also launched a free open-source code scanner.

The company says scan reports are not manually reviewed, expects a true positive rate above 90%, and acknowledges there will be false positives. Industrial equipment cannot be taken offline for updates at any time, and once vulnerabilities are found, they still need to be validated, scheduled, and fixed. What Anthropic is seeking is entry into the existing operations and maintenance service chain, not letting Claude directly take over power plants. The commercial value of AI defense depends on shortening the patching chain, and security responsibility also falls jointly on partners and facility operators. This pushes security investment from pre-release model testing to the stage of continuous monitoring and patching of production systems.

(Sources: Anthropic / Axios / Booz Allen)

4 | Axios discloses details of U.S. electricity proposal, large data centers may have to bear grid expansion costs

Axios reported on October 8 that a bipartisan U.S. Senate energy permitting plan proposes writing data center expansion and grid cost sharing into the same proposal. The plan, published on September 30, may require new data centers of more than 20 megawatts to bear the system costs of new generation, storage, and transmission and distribution. After stopping power purchases, some obligations may still continue. A 2025 update from Berkeley Lab estimates that by 2030 data centers could account for 9.5% to 15.3% of U.S. electricity use. The plan still awaits legislation, and whether household electricity bills are affected also depends on rate implementation in each locality. AI expansion has moved from cloud capital expenditure to the electricity meter: faster expansion can be discussed, but who pays for the grid cannot be avoided. If newly built computing facilities have to bear their own grid connection costs, project financing will need to package grid investment together with land and chips.

(Sources: Axios / Reuters / Lawrence Berkeley National Laboratory)

5 | Trump says no resumption of airstrikes on Iran before November 3

Trump said on Thursday that the United States will not resume airstrikes on Iran before the November 3 midterm elections, and said the two sides are holding "productive discussions." But this is not a ceasefire announcement: shipping traffic in the Strait of Hormuz is still below pre-war levels, Iran has continued to attack oil tankers recently, and U.S. escort and blockade operations are still ongoing. Brent closed on Thursday at $104.28 per barrel, up 4.1%. In early trading it was near $106, briefly fell to about $103 after Trump's remarks, and then rose again.

What the market received is a political limitation on the window for military action, not a promise of restored supply. A short-term ceasefire in June collapsed within weeks, and what happens next depends on Iran's formal response and traffic volumes through the strait. The election calendar limits public military escalation, but it will not automatically lift the blockade or restore shipping, and strait risk will still feed into oil prices and freight rates.

(Sources: Associated Press / Axios / The Economist)

Also worth knowing ↓

CrowdStrike says hackers used China-developed AI tools to attack South Korean banks. The New York Times relayed this based on the company's research; existing reporting supports the judgment that the tools were used, but this alone cannot establish that the operation was directed by the Chinese government. (Source: The New York Times)

Nvidia-backed Firmus shelves Australian IPO, considers shifting to private funding. The company cited market conditions, with the originally planned multi-billion-dollar raise shifting to as-yet-undetermined private financing; AI data center funding demand has not been realized in step with the IPO window. (Source: The Wall Street Journal / Bloomberg)

OpenAI's annualized revenue approaches $50 billion, about $20 billion lower than previously reported. This is an estimate of the current run rate, not full-year realized revenue. Different reports also use somewhat different company forecast methodologies. (Source: The Information / TechCrunch)

SpaceX plans to acquire low-band spectrum to drive Starlink mobile service expansion. The deal still requires approval from the U.S. Federal Communications Commission; obtaining spectrum does not mean nationwide commercial coverage is already in place. (Source: Bloomberg / The Verge)

Uber and Pony.ai plan to test robotaxis in London within weeks. The two companies say they will use Pony.ai's seventh-generation vehicle model; the commercial operating model and fleet ownership remain to be clarified. Uber continues to expand its platform through multiple autonomous driving companies, with vehicle technology and operational resources provided by partners. (Source: TechCrunch)

The United States suspends the eligibility of Microsoft, Adobe and other companies to participate in the PERM employment-based immigration program. The government says it is investigating employer fraud allegations; the suspension of eligibility is an administrative measure and cannot be directly taken as a ruling that the relevant companies have been found to have violated the law. (Source: Reuters / Wired)

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia