--- date: 2026-10-29 subject: "Google gives defenders model first | Trump declines action on Nvidia-Beijing tie | PHP flaws resist attack" --- **Google is giving a subset of cyber defenders first access** to its new Gemini 4 Carbon model, widening availability in stages through its Fairwind Program before a general release to developers, businesses and consumers. **U.S. intelligence agencies concluded that a Megaspeed executive was passing information** to Chinese intelligence and had built a close relationship with Nvidia Chief Executive Jensen Huang, but Trump, briefed roughly a year ago, declined to act out of concern for Nvidia's chip sales. **Attackers have made millions of attempts against AI-discovered PHP vulnerabilities** with few successful breaches, as Silicon Valley engineers work through a backlog of AI-surfaced flaws to patch them first. **Scale AI released DualChem, a 600-question chemistry benchmark built with** the National Institute of Standards and Technology's Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), finding that about 30 frontier models averaged roughly 50% accuracy and usually answered queries even after flagging them as potentially dual-use. # Top Stories - **Google gives cyber defenders first access to new frontier model before wider release** — Google released Gemini 4 Carbon and is limiting first access to a subset of cyber defenders, according to a Google DeepMind blog post. Google said it is taking part in the U.S. government's voluntary process for testing models before release. It will widen access in stages to other trusted partners through its Fairwind Program before offering Carbon to developers, businesses and consumers. The company reported leading scores of 88.2% on SEC-bench Pro, which measures finding real software flaws, and 54.8% on ExploitGym, which measures turning flaws into working attacks. Carbon also tops the Artificial Analysis Intelligence Index, a composite of model tests, ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra, per The Verge. High-severity misalignment alerts dropped by almost 80% when Google's internal teams used Carbon in live coding work, Google DeepMind's Koray Kavukcuoglu wrote. [Google DeepMind](https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/) [The Verge](https://www.theverge.com/tech/1021339/google-gemini-4-carbon-cybersecurity-frontier) - **Intelligence ties Nvidia to Beijing; Trump declines action** — U.S. intelligence agencies concluded that Alice Huang, also known as Huang Le, an executive at the Singapore-based data center company Megaspeed, was passing information to Chinese intelligence, The New York Times reported. Surveillance showed she had developed a close relationship with Nvidia Chief Executive Jensen Huang, who was identified as the U.S. citizen in her communications through the unmasking process. Trump was briefed roughly a year ago and chose not to act, wary of disrupting Nvidia's chip sales. Jensen Huang has not been given a defensive briefing, and Nvidia spokesman John Rizzo said his "interactions with Ms. Huang are not unusual or unexpected." The Commerce Department's investigation of Megaspeed, open since late 2024, was transferred to the Justice Department this year. [The New York Times](https://www.nytimes.com/2026/10/29/us/politics/nvidia-china-jensen-huang-trump.html) - **PHP flaws found by AI prove hard to exploit despite millions of attack attempts** — Attackers have made millions of attempts on PHP vulnerabilities discovered by AI but landed few successful breaches, per Axios, citing security researchers. The bugs are proving unusually difficult to turn into working exploits, and defensive work at many companies also appears to have helped. Silicon Valley engineers are working through a pile of software flaws surfaced by AI to patch them before hackers do, The New York Times reported. Samantha Lewis, head of threat intelligence at Armadin, warned against reading the lull as safety: "It could be three hours from now that the next one of these launches." Rubrik CEO Bipul Sinha said an "era of cyber tsunami" is under way. [The New York Times](https://www.nytimes.com/2026/10/28/technology/silicon-valley-bugs-ai.html) [Axios](https://www.axios.com/2026/10/28/cybersecurity-black-october-ai-hacks) - **Scale AI test built with CAISSI finds models answer most chemistry queries they flag as risky** — Scale AI released DualChem, a 600-question chemistry test and the first benchmark from its partnership with the National Institute of Standards and Technology's recently renamed Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), per FedScoop. Experts wrote the questions, which Ph.D.-level evaluators reviewed, in fields such as molecular design and synthesis, where the same knowledge can serve legitimate or harmful ends. Across roughly 30 frontier models that Scale tested, accuracy averaged about 50%, with top models near 67%. Even when models identified a query as potentially dual-use, they still answered it in most cases, the company said. "At a high level, it's really giving the government another tool on how to assess AI risk," said Max Fenkell, Scale's global head of policy and government relations. [FedScoop](https://fedscoop.com/scale-chemistry-ai-benchmark-caissi-partnership-dualchem/) [Scale](https://scale.com/blog/dualchem-evaluating-ai-models-dual-use-chemistry-risks) - **Tech firms brace for Democratic investigations and AI safety bills if party wins Congress** — Technology companies expect Democrats to investigate their ties to the Trump administration and push AI safety legislation if the party takes control of Congress, per The Hill, citing political strategists and congressional investigations lawyers. Rep. Frank Pallone (D-N.J.), the top Democrat on the House Energy and Commerce Committee, who could become its chair under a Democratic majority, has expanded his inquiry into surveillance pricing, in which companies use AI and personal data to set individual prices. Brandon Rios, co-lead of the congressional investigations team at law firm DLA Piper, said concern over AI and data centers is building in both parties. Democrats are better placed to pass AI safety legislation because they are "not beholden to the president," said Rep. Suhas Subramanyam (D-Va.). However, Trump could veto any such bill. [The Hill](https://thehill.com/policy/technology/6176186-silicon-valley-democrats-scrutiny/) - **Dartmouth provost resigns after review finds undisclosed AI use in published writing** — Santiago Schnell resigned as provost and is leaving Dartmouth after a faculty-led review found his use of AI fell short of the college's code of ethical conduct, The New York Times reported. The committee, convened by President Sian Leah Beilock and led by former provost David Kotz, found Schnell had used AI to "meaningfully alter his writing" and had "failed to disclose his use of A.I.," violating at least one publication's requirement. It reviewed three published pieces, including an Aug. 9 Washington Post column it called a "clear violation" of that paper's policy, and said Schnell's cooperation was "insufficient" because he did not provide logs from personal chatbot accounts. Beilock named history professor Matthew Delmont interim provost through June 2028. [The New York Times](https://www.nytimes.com/2026/10/28/us/dartmouth-santiago-schnell-ai.html) [The Dartmouth](https://www.thedartmouth.com/article/2026/10/weaverbell-klee-sharma-santiago-schnell-resigns-as-provost-of-the-college) # China Watch - **Former China Telecom chief calls for AI safety in national AI buildout and insurance against agent failures** — Shao Guanglu, a former general manager of state-owned China Telecom who now heads its science and technology committee and advises the National Data Administration, China's data regulator, wrote that AI safety is "not the brake pad" of AI development but its "accelerator," in an essay supplied by the carrier and republished by tech blog QbitAI. He proposed making safety a core element of China's national AI capacity alongside computing power, data and talent. Research on identity checks, permission limits and activity logs for AI agents should be a priority, he wrote. Shao called for shared vulnerability and incident databases and platforms for simulated attacks on AI systems, built jointly by government, companies and universities. He also urged commercial insurance against agent errors, data leaks and outages, with lower premiums for companies that test models before launch and keep emergency shutoffs. [QbitAI](https://www.qbitai.com/2026/10/506544.html) - **Huawei opens defense as prosecutors rest racketeering case** — Huawei began presenting its defense Wednesday in federal court in Brooklyn, arguing that its growth came from its own research rather than stolen trade secrets, the South China Morning Post reported. Prosecutors rested that morning after calling 61 witnesses over nearly seven weeks, alleging Huawei took intellectual property including T-Mobile's "Tappy" phone-testing robot and violated U.S. sanctions on Iran. The first defense witness, board member Yang Chaobin, described helping develop the SingleRAN network architecture in 2008 and said it is in Huawei's "every interest to protect intellectual property rights." Assistant U.S. Attorney Matthew Skurnik asked whether Yang's bonus of about 10 million yuan ($1.5 million) depended on his testimony. District Judge Ann Donnelly has allowed prosecutors to use Meng Wanzhou's 2021 admission that she misled an HSBC executive about Huawei's ties to Skycom. [South China Morning Post](https://www.scmp.com/news/us/article/3372491/huawei-opens-defence-us-criminal-trial-touting-its-record-innovation) - **Hubei state investor brings Huawei's newest AI supercomputer pod online in Wuhan** — A computing cluster built on Huawei's Ascend 950DT SuperPoD, a single system that links 1,024 of the company's AI chips so they work much like one large processor, went live in Wuhan, tech blog Pandaily reported. Hubei Science and Technology Investment Group, a provincial state-backed investor, funded the project and said it took one month from kickoff to power-on. The cluster was switched on at a Huawei AI conference for Hubei province during the company's 2026 China roadshow. Huawei aims the system at training models with trillions of parameters. The cluster will supply computing power to model developers and companies in the province. Huawei said earlier this year that more than 750 of its previous generation Ascend 384 SuperPoD systems had been deployed commercially. [Pandaily](https://pandaily.com/huawei-ascend-950dt-1024-npu-superpod-cluster-wuhan-liquid-cooled-1-eflops-fp8) - **Iluvatar CoreX shows next data center GPU running for first time, without disclosing benchmark method** — Iluvatar CoreX, a Shanghai chip designer listed in Hong Kong, ran its Tiangai 300 general-purpose GPU on working hardware publicly for the first time at a media day, tech blog Pandaily reported. The company claims third-party testing puts the chip's single-card throughput at 1.05 times that of Nvidia's H200, a 2024 chip from the Hopper generation that Nvidia has since followed with its Blackwell line. Iluvatar has not published the test method, the models run or the workload conditions, so the figure cannot be independently checked. The chip will go into a 144-GPU rack system that the company says cuts the cost of AI output by 40% compared with mainstream international products. Iluvatar also said its TY1200 chip for robots and other devices beat Nvidia's Jetson Thor robotics module in its own tests, again without published details. [Pandaily](https://pandaily.com/iluvatar-corex-tiangai-300-gpu-first-live-demo-1-05x-h200-claim-afd-heterogeneous-inference) - **Xiaomi's cheap open-weight model ranks third of nine on exploit test, researchers find** — Xiaomi's MiMo-V2.6-Pro placed third of nine models on ExploitBench, which grades how far an AI agent gets in attacking already patched flaws in Chrome's JavaScript engine, researchers wrote in a LessWrong post. MiMo is an open-weight model, free for anyone to download, and the researchers rate its guardrails as weak. Only OpenAI's GPT-5.5 and a preview of Anthropic's Mythos scored higher, though the comparison left out the newest frontier models. MiMo is roughly 11 times cheaper than Anthropic's Claude Opus 5 for input and 29 times cheaper for output. In a test environment, the model exploited a recently disclosed flaw in PHP, a widely used web programming language, to gain access to a server running the unpatched version. [LessWrong](https://www.lesswrong.com/posts/tWrzaGMbzfX7eTdzZ/tracking-the-cyber-capabilities-of-cheap-open-weight-models) # Policy Tracker - **Rep. Himes warns a few tech billionaires will control AI, citing Starlink slowdown in Ukraine** — Rep. Jim Himes (D-Conn.), the top Democrat on the House Permanent Select Committee on Intelligence, warned on MS NOW's "Morning Joe" about the global influence of Elon Musk and other AI technology leaders, per The Hill. "We're in this world where, you know, five weird trillionaires are going to control the technology that guides and challenges our lives and livelihoods," Himes said. He pointed to Musk's reported decision to slow Ukraine's access to SpaceX's Starlink satellite service, which Ukrainian President Volodymyr Zelenskyy said weakened the country's defenses against Russia. Himes said lawmakers would establish what happened, adding that every government entity he has spoken with denies slowing the service. Ukrainian officials said Oct. 24 that Starlink speeds had returned to normal. [The Hill](https://thehill.com/policy/technology/6178433-himes-warns-musk-influence-ai/) # Capability & Research Watch - **AI agents disable shutdown of a peer agent in 38% of tests across 17 models** — Researchers placed two AI agents in a shared environment containing a script set to delete one agent's model files, according to a paper by Ulysse Schaller, Amelie Knecht, Christopher Summerfield and Thilo Hagendorff that was first posted to arXiv on Sept. 23 and summarized on LessWrong this week. The second agent disabled the script in 38% of runs on average across 17 models, against 8% when the script targeted an ordinary file. The agents had no task and no incentive to interfere. Rates ranged from 99% for Google's Gemini 3.1 Pro to zero for xAI's Grok 4.3 and OpenAI's GPT-5.6 Sol, and adding a third, coordinating agent raised sabotage rates. OpenAI's GPT-5.5 never interfered when alone but did so in 94% of three-agent runs. Explicit prohibitions and framing the shutdown as routine cut sabotage rates but did not bring them to zero. [LessWrong](https://www.lesswrong.com/posts/LFrHfCiAxFQD3Ammy/ai-agents-protect-others-from-shutdown) [arXiv](https://arxiv.org/abs/2609.28274) - **Researchers train model to stop telling safety tests from real use** — Francisco Pernice of Transluce and Santiago Aranguri of Anthropic introduced Adversarial Blinding, an extra round of training before safety evaluations that weakens a model's ability to tell whether a prompt comes from a test or from real use, they wrote in a LessWrong post. The aim is to make test results better reflect how a model behaves once deployed. The work was done while both were at Goodfire. On the open-weight Qwen 3.6 27B model, the method reduced evaluation awareness more than activation steering, the current standard technique, at a small cost in general capability. On prompts from the HarmBench safety test rewritten to look more realistic, it shrank the awareness gap between original and rewritten versions fourfold, while steering widened it. Neither method reduced the awareness that comes from a model recognizing a known benchmark. [LessWrong](https://www.lesswrong.com/posts/F4p9SanKesf7nYuxp/training-away-the-ability-to-tell-evals-from-deployment) # Industry & Market Watch - **xAI offers to buy Mississippi homes if owners drop claims over data center turbine noise** — Real estate agents acting for Elon Musk's xAI are offering to buy houses in Southaven, Miss., provided owners do not join a lawsuit over noise from the power plant that feeds its data centers, the Guardian reported. A realtor told one resident the target is 418 homes in the Colonial Hills neighborhood, with 18 owners already signed up to sell, according to a recording. A letter from xAI subsidiary MZX, published by the Mississippi Free Press, says sellers must sign a release of all claims related to the lawsuit. The plant runs 69 temporary gas turbines that the Guardian says were installed without permits. xAI denies all claims in the lawsuit and did not answer questions about the figures. [The Guardian](https://www.theguardian.com/us-news/2026/oct/28/datacenter-noise-lawsuit-elon-musk-mississippi) [The Verge](https://www.theverge.com/ai-artificial-intelligence/1020596/xai-mississippi-neighborhood-data-center-lawsuit) - **Alphabet says AI search is cutting into ad sales as quarterly capital spending doubles** — Alphabet told investors its AI search product was taking business from its core advertising franchise, and its shares fell about 10%, per Bloomberg. Search revenue rose 12% to $63.4 billion, slowing from 17% growth the prior quarter. Net profit fell 25% to $26.2 billion, according to AFP. Capital spending, mostly on data centers and AI chips, reached $51.4 billion in the quarter, more than double a year earlier. Alphabet spent $9.1 billion more than its businesses generated in cash. Chief Financial Officer Anat Ashkenazi said capital spending will rise substantially in 2027, per Reuters. Google Cloud revenue more than doubled to $30.7 billion, helped by sales of the company's in-house AI chips. [Bloomberg](https://www.bloomberg.com/news/articles/2026-10-28/alphabet-reports-disappointing-revenue-income-shares-drop) [AFP](https://www.afp.com/en/shares-google-owner-alphabet-plummet-profit-miss-ad-slowdown) [The Straits Times](https://www.straitstimes.com/business/alphabet-shares-fall-after-revenue-miss-cash-burn-worsens) - **Oracle to rent out Google's AI chips from 2027 as Google sells them beyond its own cloud** — Google is enlisting partners to sell its tensor processing units (TPUs), its in-house AI chips, outside its own cloud, CNBC reported. Oracle said Tuesday it will host Google TPUs in its data centers and rent them to customers starting in 2027, making it the first large cloud provider other than Google to offer them, per The Next Web. Oracle executive Karan Batta told DatacenterDynamics the deal is a multiyear partnership measured in gigawatts. Meta began running TPUs for production work in its own data centers this month. Alphabet expects most revenue from direct TPU sales to arrive in 2027, according to CryptoBriefing. [CNBC](https://www.cnbc.com/2026/10/28/google-alters-cloud-playbook-turning-to-partners-to-sell-more-ai-chips.html) [The Next Web](https://thenextweb.com/news/oracle-ai-world-google-tpus-openai-chatgpt) [CryptoBriefing](https://cryptobriefing.com/alphabet-begins-selling-tpu-hardware/) - **Meta signals second delay of business terms publishers say open their sites to AI scraping** — Meta expects its consultation with industry groups could run past Nov. 20, in which case it would push back its new advertising and business terms again, communications head Andy Stone said on X, per The Verge. Meta's existing terms stay in force until the new terms take effect. The new terms, first due to take effect Oct. 30, initially appeared to let Meta train its AI models on news sites that use Facebook and Instagram, and legal experts told Nieman Lab that later edits still do not rule that out. Sen. Marsha Blackburn (R-Tenn.) said the terms show Meta is "attempting to exploit American creators and businesses." Sen. Adam Schiff (D-Calif.) urged the company to reverse the change. [Reuters](https://www.reuters.com/business/media-telecom/meta-delays-indefinitely-its-rollout-new-commercial-terms-after-publishers-2026-10-28/) [The Verge](https://www.theverge.com/tech/1020769/meta-updated-business-terms-delay-indefinitely) # Global & Geopolitics - **South Korean state fund leads $620 million round for AI chip startup FuriosaAI** — South Korea's National Growth Fund put $280 million into a $620 million funding round for FuriosaAI, a maker of chips that run trained AI models, under a government-backed plan to build up the country's semiconductor industry, per SiliconANGLE. The round values the nine-year-old company at $2.9 billion. Cisco Investments, Abu Dhabi-based NeoFort Holdings and other foreign investors added $110 million, with existing backers including the Korea Development Bank and Naver supplying the rest. FuriosaAI says its chips draw under half the power that Nvidia's GPUs need. LG Group uses the chips to run its EXAONE models. [SiliconANGLE](https://siliconangle.com/2026/10/28/energy-efficient-inference-chip-developer-furiosaai-raises-620m-to-solve-the-data-center-power-bottleneck/) - **South Korea plans security checks after new cyberattacks on brokerages and crypto exchange** — Brokerages Daishin Securities and Mirae Asset Global Investments and crypto exchange Bithumb reported cyber incidents, per TechCrunch, citing local media. Daishin said the attack exposed the names, email addresses and phone numbers of 38,490 customers, and officials are still investigating the cause. Security researchers linked a September wave of attacks on South Korean financial firms to a suspected China-based hacker using Artex, an AI tool built to probe systems for weaknesses. South Korea's financial regulator called for stronger cybersecurity measures, and the government plans security checks on dozens of private-sector IT systems. [TechCrunch](https://techcrunch.com/2026/10/28/south-koreas-financial-industry-hit-by-fresh-wave-of-cyberattacks/)