Silicon Valley Daily
MIND raised $72 million to automate data-loss prevention as security moves closer to the center of the AI business. Reports of Hacktron’s earlier intrusion into OpenAI show how capable models can help researchers reach sensitive systems. California is commissioning new safeguards, while independent evaluators and European rivals challenge how oversight should work. Capital continues to move: Manus is discussing a new round, SK hynix is widening its investment reach, and NASA awarded SpaceX three more crew flights.
Researchers used Claude to reach OpenAI employee accounts
Hacktron researchers say they used Claude to compromise OpenAI employee accounts and demonstrate repository access in July. They reported the vulnerabilities, and OpenAI says it fixed them; the researchers say they avoided reading internal code.
Newsom orders recommendations for stronger AI oversight
California Governor Gavin Newsom ordered an expert group to recommend stronger AI safeguards within two months. Independent evaluators inside labs and emergency shutoffs are among the proposals under consideration, rather than requirements already imposed by this order.
AI evaluators demand independence, access and protection from retaliation
More than 100 experts signed an AI Evaluator Forum letter setting minimum conditions for evaluators embedded in frontier labs. They want control over their findings, meaningful access to systems and staff, disclosure of conflicts, and protection against retaliation.
Times lawsuit exposes internal doubts about AI training and publishing
Newly unredacted material in the New York Times lawsuit quotes Microsoft and OpenAI employees discussing scraping and harm to publishers. Much of the reporting draws on the Times’ legal brief while underlying exhibits remain sealed, so these are litigants’ claims and selected quotations, not a court ruling.
European AI rivals challenge the proposed frontier slowdown
Reuters reports that Mistral and other European AI companies accuse US rivals of using safety concerns to reinforce their dominance. The debate also reaches customers: executives interviewed at Dreamforce told CNBC that older models already meet many business needs.
Sony Music and Universal sue Suno over its new model
Sony Music and Universal Music filed a new copyright lawsuit against Suno in Massachusetts. They allege that training its new model on outputs from earlier models carries forward infringement; Suno did not immediately comment to Variety.
Disney appoints Character.AI chief as its first CTO
Disney named Character.AI chief Karandeep Anand to its newly created chief technology officer role, effective October 2. He will oversee enterprise technology, infrastructure, data and AI platforms, product and engineering across the company.
Manus seeks $500 million at a $4 billion valuation
Manus is discussing a $500 million raise at a $4 billion valuation, TechCrunch reports, citing the Wall Street Journal’s sources. The company has resumed independent operations after its proposed Meta merger fell apart; the new financing is still under discussion.
Solidigm reportedly weighs a US NAND factory
Reuters sources say SK hynix subsidiary Solidigm is considering a NAND flash factory in the United States. The report also describes SK hynix talks with Intel about memory production in Ohio; these remain reported plans rather than completed investments.
SK hynix expands venture investing across the AI ecosystem
SK hynix launched the SK hynix Ventures brand and held its first Ventures Day in Silicon Valley. Building on an existing investment operation, it plans to back AI computing, data centers, system software and optical interconnects.
Anthropic establishes a Bay Area biology laboratory
Anthropic’s life-sciences head told Reuters the company has set up a Bay Area wet lab for physical biology work. The lab extends its AI research into drug science and disease-treatment research.
Google tests CC as a shared assistant for households
Google is testing a redesigned CC agent that helps households coordinate shared emails, calendars and tasks. The experiment supports up to six adults and is initially limited to US users with personal Gmail accounts.
Meta brings Muse’s computer actions to the Mac
Meta released Muse for Mac, where the assistant can work with files, messages, calendars, notes and mail. Meta says users choose what it can access and must approve sensitive actions.
NASA awards SpaceX three more crew flights for $946 million
NASA added three SpaceX crew missions to its space-station transportation contract for $946 million. The modification covers Crew-15 through Crew-17 and related services through 2030, bringing SpaceX’s total under the contract to 17 operational missions.
Joby reports a 3,100-mile autonomous cross-country flight
Joby says a Cessna Caravan using its autonomy system flew more than 3,100 miles across the United States without a pilot taking control. A pilot remained aboard for compliance and teams supervised remotely; this demonstrated its conventional-aircraft autonomy technology.
Joint advisory links fake job interviews to 30,000 infected devices
A multinational advisory attributes at least 30,000 compromised computers to the North Korean-linked WaterPlum campaign. Attackers posed as recruiters and distributed malware through supposed coding tests, targeting jobseekers’ credentials and cryptocurrency.
Google describes an undercover investigation inside TeamPCP
Google’s threat-intelligence team told Wired an undercover analyst entered TeamPCP’s inner circle during its software supply-chain attacks. The access helped investigators warn victims, disrupt exploitation and pass identifying information to law enforcement.
US Treasury sanctions Iran’s BitBank crypto exchange
The US Treasury sanctioned Iran’s BitBank and associated entities and individuals on September 17. Treasury alleges the exchange helped transfer bitcoin to the Islamic Revolutionary Guard Corps and process Hormuz-related payments.
House advances a bill on data-center electricity costs
The House passed the Ratepayer Protection Act on September 16 to have states consider standards making large electricity users cover related grid costs. Senate action and a presidential signature are still required; it is not yet an enacted nationwide billing rule.
OpenAI and Anthropic reportedly seek smaller compute deals
CNBC sources say OpenAI and Anthropic are exploring data-center deals around 20–30 megawatts alongside their much larger projects. Analysts cited speed to usable capacity as an attraction; the specific commercial discussions remain unconfirmed by the companies.
Xiaomi streams MiMo reinforcement-learning training metrics
Xiaomi is showing live reinforcement-learning post-training metrics for MiMo-V2.6 Pro and Flash on a public dashboard. The display includes training steps, tokens, rewards and cost, offering a view of the process rather than a finished-model evaluation.
Anthropic explains Claude’s measured role in its own research
Anthropic reports that Claude led 26% of measured AI research-and-development work in August under its automation scale. That level still includes human supervision, and none of the measured work reached full autonomy; the ratings themselves partly relied on Claude.
OpenAI introduces Astra for Law with dedicated legal search
OpenAI introduced Astra for Law, combining GPT-6 Astra with a legal search index and instructions for legal analysis and writing. The index covers US case law, statutes and regulations, targeting professional research workflows.
UN and Google launch a searchable global-statistics platform
The UN system launched Data Commons, an open platform built on Google’s technology that connects statistics across agencies. Natural-language search and shared geographic and time references aim to reduce the work needed to combine datasets.
Mazama raises $135 million for superhot-rock geothermal development
Mazama Energy says it raised a $135 million Series B to develop wells in superhot rock. The company plans to begin generating electricity next year; its proposed per-well output and wider site capacity remain development targets.
DeepMind launches a forum for competing views on AGI
Google DeepMind launched an institute intended to broaden discussion of advanced AI and its social implications. It aims to present differing views from Google, DeepMind and the wider research community, including disagreements within that group.
Google expands its AI-and-economy research team
Google is expanding its AI & Economy program with academic advisers, visiting fellows and research leaders. Philippe Aghion and Ajay Agrawal are among the additions as the group studies AI adoption, work and productivity.
YC reports more solo founders and a shift toward hard tech
Y Combinator says nearly one in five companies in its cohort has a solo founder and describes more startups tackling physical-world problems. Its Lightcone discussion attributes some of that shift to smaller teams using AI; these are observations about YC’s own sample, not the entire startup market.
Fal describes the engineering behind real-time generative video
In an a16z interview, fal’s team describes combining post-training of MiniMax’s open-weight video model with systems and hardware optimization. They argue that professional tools also need predictable control over cameras, lighting, characters and motion, beyond faster generation.
Apple’s new iPhone, Watch and AirPods arrive in stores
Apple says the iPhone 18 Pro lineup arrived in its stores worldwide on September 18. Apple Watch Series 12, Apple Watch Ultra 4 and AirPods 5 joined the same retail launch.
CXMT reportedly explores a move into NAND flash
Tom’s Hardware reports that DRAM maker CXMT is preparing research and pilot-production plans for 3D NAND near Beijing. The reported work would extend the company into flash storage, with the plans still described as unconfirmed industry reporting.