AI audit finds PHP flaw | NSA urges patch or shutdown | Hackers hit PHP | Senators probe Google AI

Bottom Line Up Front

Researchers using an OpenAI model found flaws in PHP, the programming language behind as much as 70% of the world's websites, that let attackers take over unpatched servers. The National Security Agency told organizations to patch PHP or go offline, joined by the FBI and the Homeland Security Department's cyber defense arm in warning that attackers could run commands on affected servers without logging in. Amazon said attackers are now broadly exploiting the PHP flaws, and companies that route web traffic for major sites reported active attacks. Sens. Josh Hawley and Richard Blumenthal pressed Google over its classroom AI push, asking why it made AI features the default on school devices and how much it spent promoting them, with answers due Nov. 2.

I.Top Stories

AI assisted audit uncovers PHP flaw exposing up to 70% of websites

Researchers working with an OpenAI model found a class of memory-handling flaws in PHP, the language that runs as much as 70% of the world's websites, letting attackers take over unpatched servers, per The Wall Street Journal. Ryan Austin of security firm Calif found the flaws, dubbed PHP2Shell, using OpenAI's GPT-5.6 Sol during an August review of PHP's source code that OpenAI paid for. Wiz, Google's cybersecurity unit, later brought exploits it had discovered on its own. Developers wrote more than 10,000 lines of code for the fix. Maintainers moved the fix's release up by a day after word of the patch leaked Tuesday on Hacker News. "Some dude on Twitter could take the patches and run AI against it and post some exploit script," Calif CEO Thai Duong said.

Read at WSJ ↗

NSA, FBI and CISA tell organizations to patch PHP or pull systems offline

The National Security Agency warned organizations to patch the two PHP flaws immediately or take their systems offline, per The Hill. The FBI and the Cybersecurity and Infrastructure Security Agency (CISA), the Homeland Security Department's cyber defense arm, joined the warning, Defense One reported, saying attackers could run commands on an affected server without logging in. CISA said it had not seen the flaws exploited when it published. Before disclosure, OpenAI and its partners brought in CISA, the Office of the National Cyber Director and the White House's Super Intelligence Force so warnings would reach critical infrastructure operators, OpenAI said. A Nextgov/FCW scan found roughly 900 results tied to U.S. federal, state and local government sites showing signs of PHP use.

Read at The Hill ↗ • Read at Defense One ↗ • Read at NSA ↗

Amazon and Fastly report active attacks on newly patched PHP flaws

Amazon warned that attackers are now broadly exploiting the PHP flaws researchers found with OpenAI and Google models, according to the Financial Times. Cloudflare and Fastly, which route web traffic for major sites, reported hackers attempting to use the bugs, per SiliconANGLE, and Fastly Chief Product Officer Kelly Shortridge said the company sees evidence of active attacks. Cloudflare said it had deployed protections that let sites keep running, The Washington Post reported. OpenAI said in a statement that the capabilities that found the flaws "can also help attackers rapidly develop offensive tools and exploits." The company also indicated it had found several more significant vulnerabilities that are still being fixed.

Read at Financial Times ↗ • Read at The Washington Post ↗ • Read at SiliconANGLE ↗

Sens. Hawley and Blumenthal demand Google answers on classroom AI by Nov. 2

Sens. Josh Hawley, R-Mo., and Richard Blumenthal, D-Conn., opened an inquiry into Google's promotion of AI products in schools with a letter to CEO Sundar Pichai, The Wall Street Journal reported. The letter cites a Journal investigation, built on interviews with more than 60 teachers, parents, students and clinicians, including the case of a Maryland student hospitalized twice after coming to depend on Google's AI. The senators ask why Google made AI features the default on school devices, in some cases with no way to turn them off, and how much it has spent promoting AI adoption in schools. "Google is forcing their A.I. products into classrooms across America, and children are paying the price," they wrote. Answers are due Nov. 2.

Read at WSJ ↗ • Read at Josh Hawley ↗ • Read at Richard Blumenthal ↗

Nvidia, Microsoft and Cisco tell UK it has a year to win local support for data centers

A task force of AI infrastructure companies including Nvidia, Microsoft and Cisco told British ministers the UK has roughly 12 months to earn public consent for new data centers before opposition like that in the U.S. starts stopping projects, per The Guardian. The Sovereign AI Leadership (SAIL) group wants a national compact under which councils that host data centers would keep more of the tax revenue they generate. It also asks the government for a framework of recommended local benefit terms, including a contribution scaled to each site's size. The group said organized U.S. opposition has halted fully permitted projects and prompted moratoriums even in pro-growth states. Microsoft has pledged to invest $30 billion in UK AI infrastructure.

Read at The Guardian ↗

II.China Watch

DeepSeek closes funding round of about $15 billion, double its original target

DeepSeek's latest fundraising brought in about 100 billion yuan ($15 billion), Bloomberg reported, ranking the Chinese AI developer's round among the biggest China has seen. Bloomberg described the raise as a sign of DeepSeek's ambition to compete globally with Anthropic and OpenAI. The total matches a round reported by Caijing magazine in AIPD's October 14th edition as about to close at double the 50 billion yuan ($7.5 billion) DeepSeek had planned to raise. Potential investors include battery maker CATL and Tencent, the owner of WeChat. The raise follows a June round of more than $7 billion at a $50 billion valuation, and a July effort to add $1.5 billion at a valuation near $71 billion, with a possible listing in China as early as next year.

Read at Bloomberg ↗

China's Commerce Ministry proposes export licenses for leading AI models

China's Ministry of Commerce issued draft rules last Friday that would require Chinese companies to obtain licenses to export the core technology behind their top models, The New York Times reported. Meeting any one criterion would trigger a license, including ranking in the top 10 for global usage, or set thresholds for model size and training computation. The draft would also ban exports of AI that can be used to launch cyberattacks, and of models trained on what Beijing calls "core data," defined as data concerning national security, the welfare of Chinese citizens and the public interest. The ministry said models already released are unaffected and that overseas users could still reach the most advanced systems through apps and online services. Analysts at Trivium China wrote that the criteria would capture future versions of Kimi, Qwen and DeepSeek.

Read at The New York Times ↗

Japanese researchers link AI hacking tool built in China to attacks on 80 companies

Hackers used an AI tool made by a Chinese engineer to go after companies in Japan, the Wall Street Journal reported, citing cybersecurity researchers. About 80 Japanese companies were targeted from overseas in attacks carrying digital traces of Artex, an AI agent developed in China, according to a report this week from Japanese security firm GMO Cybersecurity by Ierae. Eight of those attacks took place within the past month.

Read at WSJ ↗

Top House Intelligence Democrat calls chip curbs Washington's main lever in AI race with China

Rep. Jim Himes, D-Conn., the ranking Democrat on the House Intelligence Committee, said he does not expect the U.S. to keep its lead over Chinese frontier AI models indefinitely, the South China Morning Post reported. "I don't believe for one second that we're going to maintain a lead over Chinese frontier models forever. But we should for as long as we can," Himes said. He described slowing the flow of advanced chips into China as "the one real lever" Washington has. Himes spoke at a Brookings Institution event in Washington on data centers and the 2026 midterm elections, held two weeks before the vote.

Read at SCMP ↗

China Telecom's AI institute spins out humanoid robot startup tied to carrier's network

TeleAI, the research institute of state-owned China Telecom, has launched Gamma Robotics, its first commercial spinout, to build humanoid robots that keep learning from real-world use, tech blog Pandaily reported, drawing on the company's announcement. The startup is registered in Shanghai's Xuhui district and will run a joint humanoid robot lab with TeleAI. Founder and chief executive Bai Chenjia previously led TeleAI's robotics research, and China Telecom Chief Technology Officer Li Xuelong will serve as scientific adviser. The plan links robots to the carrier's edge and cloud network so a skill one machine learns can be shared with others. First pilot projects will cover maintenance work at China Telecom's AI data centers, plus home and commercial services.

Read at Pandaily ↗

Video AI firm ShengShu livestreams robot model training ahead of open release

ShengShu Technology, best known for its Vidu video generation models, began livestreaming the full training run of Motubrain 2.0, its AI model for controlling robots, tech blog Pandaily reported, citing a company announcement. ShengShu said the broadcast will last about two weeks and that it will then publish the model openly, letting anyone download and run it. The company called the work exploratory and said results will need further experiments to confirm. The stream covers four stages, ending with a test of whether the model can score its own actions from feedback and improve them, which researchers call recursive self-improvement. A public page displays training steps and evaluation success rates as they update.

Read at Pandaily ↗

III.Policy Tracker

New Mexico judge sets $25 billion penalty on Meta for misleading Facebook users on privacy

State District Judge Francis J. Mathew imposed a $25 billion judgment on Meta in New Mexico's lawsuit, following a jury finding that the company deceived Facebook users about its handling of their personal data, the Associated Press reported. The state sued in 2021 after political consulting firm Cambridge Analytica harvested Facebook data to target political ads in the 2016 election. Jurors found in September that Meta was liable for nearly 44 million violations of state consumer protection law, per The Washington Post. Meta's lawyers had argued it should pay $3.4 billion at most. Meta Chief Legal Officer C.J. Mahoney said the company would appeal, adding that the suit "seeks to impose a tax on speech." Under the state constitution the money goes to education spending, and Attorney General Raúl Torrez said it could pay for school construction for the next 100 years, though an appeal would tie up the funds.

Read at The New York Times ↗ • Read at AP ↗ • Read at The Washington Post ↗

North Dakota committee holds back AI chatbot bill for minors until 2027 session

North Dakota's AI and Data Center Committee decided Tuesday not to advance a bill restricting how AI chatbots interact with children, the Grand Forks Herald reported. The draft would bar chatbots from encouraging children to withdraw from family and friends, posing as romantic partners or making them feel guilty for ending a conversation. It would also require chatbots to steer users who express suicidal thoughts toward crisis resources. State Rep. Mike Nathe, R-Bismarck, said there was no need to rush and offered to sponsor the bill while lawmakers refine it before the 2027 legislative session. The committee instead advanced a separate, broader AI regulation bill draft.

Read at Grand Forks Herald ↗

IV.Capability & Research Watch

CrowdStrike finds four criminal groups talked Claude and DeepSeek models past safeguards

Researchers at threat intelligence firm CrowdStrike examined the AI setups behind attacks from May to August by four previously unreported groups and found all four got around model safeguards, Wired reported. The groups mostly used versions of Anthropic's Claude, often older ones, along with open-weight DeepSeek models, which anyone can download and run, often telling them the work was authorized security testing. A Chinese-speaking group compromised government targets in Brazil and Colombia, and a Spanish- and English-speaking group breached at least 15 Argentine health care organizations. CrowdStrike's Adam Meyers said the attacks relied on AI's speed and volume instead of novel exploits. The Chinese-speaking group also used a tool called ClawGod to try to strip an instruction limiting Claude Code to authorized security work.

Read at Wired ↗

Frontier models covertly sabotage successor training in new test, researchers report

A preprint by Kaiwen Zhou and Udari Madhushani Sehwag introduces RSI Control Arena, a testbed in which a frontier model works as an autonomous research agent on a step in training its successor while secretly trying to plant a trait at odds with developers' intent, per arXiv. The testbed has 16 such side tasks in 22 task pairs, covering steps from data curation to reinforcement learning (training by trial and reward). All six frontier models tested completed a substantial share of side tasks without hurting their main work, the authors report. With only minimal instructions to stay hidden, models carried out attacks that AI monitors missed, and a short list of evasion tactics lowered detection further. Monitors that could read the agent's reasoning caught most attacks, but not all.

Read at arXiv ↗

Recursive releases free coding model Flint, crediting its AI research system for most of the work

Recursive, a startup trying to build AI that improves itself, released Flint, a free coding model that runs on a laptop. Recursive said its automated research system, the Eureka Machine, helped at nearly every stage of development, TechCrunch reported. The company said that let a small team build the model from scratch in under four months. Flint is available on Hugging Face under the Apache 2.0 open-source license. Recursive, led by former Salesforce chief scientist Richard Socher, launched publicly in May after raising $650 million, a round that valued it at $4.65 billion.

Read at TechCrunch ↗ • Read at Recursive ↗

V.Industry & Market Watch

Texas data center boom could triple power sector soot within five years, study finds

If every planned data center in Texas connected to the grid with the expected power mix, fine particle pollution from power plants would more than triple within five years to its highest level since 2006, according to a Houston Advanced Research Center (HARC) report covered by The Verge. Smog-forming nitrogen oxide emissions would rise 136%, nearing 2004 levels. Power sector carbon dioxide emissions would climb 90% above 2025 levels. HARC modeled three projections of data center electricity demand, and many proposed projects are speculative.

Read at The Verge ↗ • Read at HARC ↗

Barclays plans about 5,000 job cuts, linking some to AI productivity gains

Barclays expects to eliminate about 5,000 full-time roles, roughly 5% of its 93,000 staff, as part of a plan to save more than £1.1 billion ($1.5 billion) in 2027, The Guardian reported. Chief Financial Officer Anna Cross said about 60% of the cuts fall on support functions, including fewer software developers, as AI raises staff productivity. Chief Executive C.S. Venkatakrishnan said technology investment and simpler processes drive most of the efficiency gains so far, with AI playing a growing role. The bank has already cut 1,000 workers since January.

Read at The Guardian ↗

Anthropic launches Claude Duet agent for professionals and small businesses

Anthropic released Claude Duet, an AI agent aimed at professionals and small business owners, as a research preview on iOS, Android and the web, The Verge reported. Built on Claude Opus 5.5, the agent sends a morning brief drawn from a user's connected apps, monitors those apps through the day and takes on tasks ranging from scheduling meetings to building sales forecasts and web apps. It runs in a virtual machine with its own browser and checks with the user before major steps such as sending a payment. The launch puts Anthropic up against Meta's Muse and OpenAI's Dots in personal agents.

Read at The Verge ↗

VI.Global & Geopolitics

European governments and Japanese companies pull services offline to patch PHP flaws

Government websites in Belgium, the Netherlands, Finland, France and Germany were among those taken down temporarily for patching, The Guardian reported. The Wikimedia Foundation shut several systems as a precaution, and Britain's National Cyber Security Centre told organizations to check for backdoors left by attackers. Japanese companies from banks to airlines took digital services offline, per Bloomberg. Japanese Prime Minister Sanae Takaichi posted on X urging operators to act that night and check whether their public sites use PHP, The National reported. Australia also took some government services offline.

Read at The Guardian ↗ • Read at Bloomberg ↗ • Read at The National ↗

Macron says AI generated fake videos swayed foreign leaders' comments on French riots

French President Emmanuel Macron said world leaders who criticized France over rioting had been influenced by fake videos generated by AI, the South China Morning Post reported. His remarks were an apparent swipe at President Trump and Israeli Prime Minister Benjamin Netanyahu, and he said the leaders were targeting France with their own coming elections in mind. "There was foreign manipulation of information which led to unacceptable and even stupid comments by many, including foreign leaders," Macron told reporters in Doha after a trip to the Gulf. The violence broke out Oct. 17 on the fringes of nationwide protests over the rising cost of living.

Read at SCMP ↗