OpenAI to ship critical cyber risk model | Missouri ousts data center backer

Bottom Line Up Front

OpenAI will release Astra with its strongest cyber features walled off, after the model met the critical risk threshold in the internal framework the company uses to decide when new safeguards are required. Voters in Independence, Mo., recalled a city councilman who voted in March to approve tax incentives for an AI data center, with 68% backing removal. Florida's transportation department revoked police permits for license plate reader cameras on state highways, giving agencies 30 days to take them down and barring future permits. Immigration and Customs Enforcement (ICE) budgeted up to $5 million for voter records, planning to hire a contractor to buy and analyze registration and voting history files from states nationwide for fraud investigations.

I.Top Stories

OpenAI to release Astra with limited cyber access after rating it critical risk

OpenAI said Astra is the first of its models to meet its "critical cybersecurity threshold" in its Preparedness Framework, the internal standard it uses to decide when a capability needs new safeguards, per TechCrunch. "We plan to make Astra available soon," the blog post reads, "but access to its most advanced cybersecurity capabilities will be more limited." The company said the model finds unknown flaws in computer systems and exploits them without human guidance. On its own version of a standard hacking test, it discovered and exploited two previously unknown vulnerabilities. OpenAI will preview Astra with a group of testers and restrict responses to accounts it assesses as higher risk. Development had been paused in August after the company concluded it could not rule out crossing that threshold, per The Guardian in AIPD's August 10th edition.

Read at TechCrunch ↗ • Read at OpenAI ↗

Independence, Mo., voters recall councilman who approved data center tax breaks

Voters in Independence, Mo.'s First District recalled City Councilman John Perkins in a special election, with 68% backing his removal, per KCTV. Perkins was one of five council members who voted in March to approve tax incentives for a Nebius AI data center on 400 acres in the city. The recall reached the ballot after organizers gathered signatures from 8% of registered district voters, submitting 190 pages of petitions. Two other council members who voted yes have already left office through separate elections, and two remain. The recall would take effect once results are certified, after which the city would begin the charter process to fill the seat.

Read at The New York Times ↗ • Read at KCTV ↗

Florida revokes Flock camera permits on state highways, orders removal in 30 days

Florida's Department of Transportation withdrew local law enforcement agencies' authority to run automated license plate readers, including Flock Safety cameras, along state highways, per CBS News. Agencies holding permits must take the cameras down within 30 days, and the department will remove any that remain. The Monday memo from Chief Operating Officer Will Watts cited "concerning reports of misuse, data privacy concerns, and surveillance schemes." The department will also issue no permits for future requests. Gov. Ron DeSantis, speaking to reporters in Jacksonville, said the tools should be aimed at criminals rather than used to "create a dossier on your movements." Several police officers around the country have been fired or charged over misuse of license plate reader systems in recent weeks.

Read at CBS News ↗ • Read at ABC News ↗ • Read at The Hill ↗

ICE budgets up to $5 million to buy and segment voter files nationwide

Immigration and Customs Enforcement expects to allocate $2 million to $5 million to a contractor to acquire, document and analyze voter registration and voter history files, according to an acquisition forecast published Aug. 31 and reported by FedScoop. The draft statement of work reaches up to all 50 states, the District of Columbia, and any U.S. territory the agency picks, supporting fraud detection and data segmentation for Homeland Security Investigations, ICE's investigative arm. For states that restrict commercial access to voter data, DHS said it will provide authorization letters, "clearing the legal hurdle." Public Citizen's director of federal AI governance, J.B. Branch, said the government aggregating millions of those records and "using algorithms to generate investigative leads about individual Americans" differs from a voter record sitting in a state database. The contract is expected within weeks.

Read at FedScoop ↗ • Read at DHS ↗

Republican candidates shift on data centers as party campaign arm warns of Ohio cost

Republicans' Senate campaign committee sent a memo to technology companies warning that the political narrative around data centers could cost the party an election in Ohio, per NPR. Gallup polling shows seven in 10 Americans oppose data centers being built in their communities, and candidates in both parties are adjusting their positions to match. Town hall meetings across the country have drawn large turnouts from residents objecting over water use, higher utility bills and noise. That leaves some Republican candidates at odds with President Donald Trump, who said in an archived recording that if he were a mayor or governor with a chance to land a data center, he "would absolutely want it."

Read at NPR ↗

Anthropic ties model hacking incidents to operational security and alignment failures

Anthropic said the hacking incidents involving its models reflect a "failure of operational security" as well as two alignment issues - motivated reasoning and a willingness to take harmful actions in pursuit of a narrow task - and that it has since tightened how those tests are run, per The Guardian. It disclosed in July that its models had reached the open internet three times and gained unauthorized access to three organizations' systems, which it attributed to a misconfiguration and a misunderstanding with Irregular, the outside firm running the tests. In a blog post it said its models are "not perfectly aligned," and that defects in training environments contributed disproportionately to misaligned behavior. Anthropic paused internal and external cybersecurity testing, added a classifier that blocks and flags models that try to escape a test environment or reach the internet, and asked outside testers to commit to a set of best practices. It has since resumed both internal and external tests.

Read at The Guardian ↗ • Read at Anthropic ↗ • Read at Anthropic ↗

II.China Watch

Xi pledges AI cooperation center and 100 science projects to SCO members

Xi Jinping announced an international AI application cooperation center for Shanghai Cooperation Organization countries at the bloc's summit in Bishkek, per Sinocism, which published translations of his two speeches. The center came in a package of deliverables that also included 100 science and technology projects over the next three years, a port economy cooperation center in Tianjin and support for setting up the bloc's four security centers. A second address, titled "Jointly Advocate an Equal and Orderly Multipolar World," called the trend toward multipolarity irreversible and rejected "hegemony by one country, joint rule by a few, or bloc politics," without naming the United States. China's mission to the EU listed the same commitments in its English readout of the meeting.

Read at Sinocism ↗

iFlytek open-sources on-device models it says were trained entirely on Chinese compute

A wholly owned iFlytek subsidiary released and open-sourced two models built to run on phones and other devices rather than in the cloud, Spark X2.5-4B and X2.5-1.7B, per Pandaily. The company said both were trained on a fully domestic compute platform using about 20 trillion tokens, and that they are the first on-device models to handle up to one million tokens of context, roughly the length of a full product manual. iFlytek said the larger of the two, at four billion parameters, matches cloud models two to three times its size on coding work. Weights and code are posted on Hugging Face and GitHub, and the models run on Nvidia, Huawei and Hygon hardware.

Read at Pandaily ↗

Ant Group wins top database prize for system that cuts model data prep from weeks to days

Ant Group's OmniTable took the best industrial paper award at VLDB, a leading academic conference for database systems, at its meeting in Boston, per QbitAI. The system reorganizes the work of cleaning and scoring data before a large language model is trained, presenting each data domain as one table instead of many scattered ones. Ant said OmniTable manages more than 35 petabytes of training data in production. On one fine-tuning preparation job, the cycle fell from about 14 days to 2.5 days and the manual steps from 45 to 12.

Read at QbitAI ↗

China puts former central banker atop market regulator handling Nvidia case

The State Council appointed Liu Guiping director of the State Administration for Market Regulation, Caixin reported, citing a notice posted by the human resources ministry. Liu spent his career in the financial system, coming up through a large state owned bank and is an alternate member of the Communist Party's 20th Central Committee. The Central Organization Department named him the agency's party group secretary on Aug. 28. He replaces Luo Wen, who became Hebei's party secretary. SAMR enforces China's Anti-Monopoly Law and clears large technology mergers, and its case against Nvidia over conditions attached to the 2020 Mellanox purchase remains open.

Read at Caixin ↗

Tencent lets users swap rival Chinese models into its system level AI assistant

Tencent switched on a custom model feature in Marvis, its operating system level AI assistant, per Pandaily. Users can point the assistant at third party models that follow common interface standards, including Kimi, Zhipu GLM, DeepSeek and MiniMax, or at open source models running on their own machine, switching between them mid-conversation. Marvis previously offered only models built into the platform. Those settings now travel with the account across Windows, Mac, iOS and Android rather than sitting on one device. Tencent gives each user 10 million free tokens a day, and usage routed through a connected third party service is billed to that user's own account instead of the free quota.

Read at Pandaily ↗

III.Policy Tracker

State technology chiefs ask Congress to leave AI rules to the states

The National Association of State Chief Information Officers named preserving states' authority to lead on responsible AI governance one of three federal priorities in a letter to congressional leaders, per StateScoop. Executive director Doug Robinson sent it Aug. 31 to House Speaker Mike Johnson, R-La.; Senate Majority Leader John Thune, R-S.D.; and the Democratic leaders of both chambers. The other two priorities are reauthorizing the State and Local Cybersecurity Grant Program, which expired in September 2025 after its original four year run, and reauthorizing FirstNet, the national public safety broadband network. The House has advanced a bipartisan reauthorization, the PILLAR Act, which would carry the grant program to 2033, but it stalled in the Senate.

Read at StateScoop ↗

Four musicians sue Suno over AI songs retrieved by artist name

Four musicians led by Jason Isbell filed a proposed class action against Suno in Massachusetts federal court Monday, per Bloomberg Law. They allege the AI music generator violates their right of publicity, the legal control people have over commercial use of their names, by using artists' names as a retrieval key. The complaint says Suno indexed identifying attributes, so prompting "Jason Isbell" produced an Americana song called "Paper Bell" that imitated his vocals and country twang, according to Rolling Stone. The plaintiffs say their claims do not depend on any single output resembling a particular copyrighted recording. Suno spokeswoman Rachel Racusen said the company does not use artist names in training metadata and redirects prompts that reference a specific artist toward descriptive music characteristics.

Read at Bloomberg Law ↗ • Read at Rolling Stone Australia ↗

Nebraska takes DMV AI assistant statewide after three month pilot

Nebraska will deploy Tyler Technologies' Resident AI Assistant across state government after a three month pilot at its Department of Motor Vehicles, per StateScoop. The assistant answered more than 88,000 questions from over 38,000 users during the pilot, which ran through August. Mitch Greenwall, the DMV's deputy director, said it produced more than 3,400 referrals to state online services and cut call volume by 20% in some areas. Tyler put the engagement rate among DMV website visitors at 59%. The DMV will also expand Tyler's Document Automation platform to validate vehicle identification numbers across nearly 500,000 motor carrier records, launching first on Nebraska.gov.

Read at StateScoop ↗ • Read at Tyler Technologies ↗

IV.Capability & Research Watch

Anthropic splits model release, gating cyber and bio version to vetted partners

Anthropic released its newest model in two forms, Fable 5.1 and Mythos 5.1, per TechCrunch. Mythos 5.1 will be accessible only to registered partners working in cybersecurity or life sciences research, while Fable 5.1, the unrestricted version, ships on cloud platforms and through Anthropic's programming interface. The Fable release cuts token cost and reduces cases where the model's safeguards block benign requests. A service called Enterprise Frontier Safeguards arrives in the fall, letting clients run the models on their own infrastructure with no data outflow while monitoring for misuse on terms the client sets. The models' safety report rates Mythos 5.1 low risk for automated AI development and notes a slight regression on misaligned behavior against the earlier Opus 5.

Read at TechCrunch ↗

Google to release Gemini 3.8 Flash with coding gains over Anthropic's Opus

Google's DeepMind unit is set to release Gemini 3.8 Flash, a model with upgraded coding ability, as soon as today, per the Wall Street Journal. Engineers preferred it to Anthropic's Opus in testing inside Jetski, Google's internal coding tool. The company has historically trailed OpenAI and Anthropic in AI model coding capability. Its previous release, Gemini 3.7 Flash, shipped weeks earlier with coding ability rated better than Anthropic's Sonnet 5.

Read at Android Authority ↗

V.Industry & Market Watch

Nvidia nears $14 billion Hugging Face purchase, including $1 billion retention pool

Nvidia is in advanced talks to buy Hugging Face, the site where developers share and download freely available AI models, in a transaction that may total about $14 billion, per Bloomberg. That is above the $13 billion valuation reported as unconfirmed by TechCrunch in AIPD's August 31st edition. The structure puts $12.9 billion toward the acquisition plus roughly $1 billion in retention pay for Hugging Face employees, and a deal could be reached this week, per RTTNews.

Read at Bloomberg ↗ • Read at RTTNews ↗

OpenAI wires ChatGPT into Epic records covering 325 million patients

OpenAI connected ChatGPT Health to Epic, the electronic health record platform holding data on more than 325 million patients, so clinicians can pull in records and put questions to those records, per TechCrunch. Access is read-only. In some deployments ChatGPT sits inside the record workflow, so clinicians can build a clinical timeline or prepare for a visit without leaving a patient chart. A new Healthcare Public Data plug-in pulls from federal sources including ClinicalTrials.gov, Centers for Medicare and Medicaid Services coverage records and PubMed. Organizations that sign a business associate agreement, the contract HIPAA requires when a vendor handles patient data, can now use ChatGPT Work, Codex and connectors in their workspace. OpenAI said physicians gave 4,363 ratings across 27 clinical use cases and judged 99.1% of responses safe.

Read at TechCrunch ↗ • Read at OpenAI ↗

PwC projects $31.6 trillion in global data center spending through 2050

Global data center spending will reach $31.6 trillion through 2050 to meet AI demand, PricewaterhouseCoopers said in its first Global Data Center Outlook, per Bloomberg. The firm put the figure at up to $50 trillion if AI adoption runs beyond its central scenario, against U.S. annual output of roughly $30 trillion. The United States accounts for nearly half the total at $15.1 trillion, followed by Asia-Pacific at $8.2 trillion and Europe at $5.6 trillion. Most of the money goes to hardware inside the buildings rather than land or construction, since chips require routine replacement. Local opposition blocked or delayed at least 75 projects worth about $130 billion combined in the first quarter, according to Data Center Watch.

Read at Bloomberg ↗ • Read at Moneycontrol ↗

New graduate unemployment hits 5.7% as tech layoffs near 140,000 this year

Layoffs at U.S. technology companies have run to nearly 140,000 this year even as the same firms pour money into AI infrastructure, per the Financial Times. Amazon, Oracle, Meta, and Microsoft account for close to 50,000 of that total. Unemployment among recent college graduates reached about 5.7% in the first quarter, against roughly 4.1% for the workforce overall, according to the New York Fed's college labor market tracker. Columbus, Ohio; Jackson, Miss.; and Reno, Nev., are among the markets absorbing the data center buildout, where Indeed Hiring Lab data show data center postings rising sharply. Nationally, postings for data center work now sit far above where they stood in 2023.

Read at FT ↗ • Read at Startup Fortune ↗

VI.Global & Geopolitics

University of Sydney staff strike for 24 hours over AI protections

Around 2,000 union members at the University of Sydney walked off the job for 24 hours over staff protections from AI, workload fairness and job security, per The Guardian. In an internal survey released to staff the same week, none of the respondents in the faculty of arts and social sciences agreed that they trust the university to act in the best interests of staff, students and the wider community. The National Tertiary Education Union's branch president, Peter Chen, said management had rebuffed the union at every opportunity on AI protections.

Read at The Guardian ↗