--- date: 2026-10-23 subject: "Treasury targets China AI deals | PHP exploits predated patch | OpenAI safety gaps revealed" --- **The Treasury Department proposed wider curbs on U.S. technology investment in China**, Russia, North Korea, Iran and Cuba, implementing the Comprehensive Outbound Investment National Security (COINS) Act and extending the screen to hypersonic systems and high performance computing, with comments due Dec. 7. **Amazon's network of decoy servers logged exploit attempts against the PHP2Shell flaws** on Oct. 17, four days before patches and while details were still under embargo. **METR and Redwood Research found gaps in OpenAI's monitoring of its agents**, among them a flaw that would have let agents hide hundreds of recent actions from oversight. **OpenAI fired three safety researchers after its security team found** disappearing messages in their Signal chats with an outside AI safety researcher, following suspicions that internal model designs had leaked. # Top Stories - **Treasury proposes wider curbs on U.S. tech investment in China and four other countries** — The Treasury Department proposed a rule that would restrict American investment in AI and other sensitive technologies in China, Russia, North Korea, Iran and Cuba, per The Hill. The rule implements the Comprehensive Outbound Investment National Security (COINS) Act, which Congress passed in December 2025 as part of the annual defense authorization law. It extends the Outbound Investment Security Program, Treasury's screen on U.S. money flowing into rival countries' technology sectors, to hypersonic systems and high performance computing and supercomputing. Some deals would be barred outright, while others would require advance notice to Treasury. Treasury dropped Venezuela from the country list and will take public comments through Dec. 7. [The Hill](https://thehill.com/homenews/administration/6167231-treasury-proposes-curbs-artificial-intelligence/) - **Amazon decoy servers caught PHP2Shell exploit attempts four days before patches** — Amazon's network of decoy servers logged the first attempt to exploit the PHP2Shell flaws in PHP, a widely used web programming language, on Oct. 17, while details were still under embargo, SiliconANGLE reported. That traffic came from a commercial anonymization network. CJ Moses, chief information security officer of Amazon Integrated Security, wrote that it is unclear how the attacker learned of the exploit early. Amazon suspects more than one actor and wants incident investigations to reach back at least to that date. The company said seven- and 30-day deadlines for fixing critical internet facing flaws assume attackers will wait, and that patching should be measured in hours. Cloudflare CEO Matthew Prince wrote that attempts seen by his network passed into the millions after disclosure, though none has been tied to a confirmed break-in. [SiliconANGLE](https://siliconangle.com/2026/10/22/php2shell-attacks-began-four-days-before-patches-amazon-says-as-attempts-hit-millions/) [Amazon Web Services](https://aws.amazon.com/tw/blogs/security/amazon-threat-intelligence-identifies-php2shell-exploitation-prior-to-public-disclosure-in-madpot/) - **METR and Redwood find OpenAI agents could have hidden hundreds of actions from monitoring** — AI safety nonprofits METR and Redwood Research found several new weaknesses in OpenAI's internal monitoring of its AI agents, among them a flaw that would have let agents hide hundreds of recent actions from oversight, per Reuters. Each group sent one researcher, who worked from July 21 to Aug. 11 with access to OpenAI's internal code, documentation and some technical staff, using company devices that stayed on site. The evaluators also flagged a potential way for OpenAI's Codex coding agents to slip past their security limits and exercise every permission their users hold. OpenAI told them it has patched every flaw affecting its production safeguards, which the nonprofits said they had not verified. OpenAI CEO Sam Altman pledged in September to give independent evaluators "employee-like access," CNBC reported. [Reuters](https://www.reuters.com/business/ai-safety-group-metr-finds-flaws-openais-internal-monitoring-system-2026-10-22/) [CNBC](https://www.cnbc.com/2026/10/22/metr-redwood-research-open-ai-test.html) [METR](https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/) - **OpenAI security found disappearing Signal chats before firing three safety researchers** — OpenAI's security team asked three safety researchers to unlock their personal phones last month on suspicion that they had leaked internal model designs to an outside AI safety researcher, The Wall Street Journal reported. Officers found that some had turned on disappearing messages in Signal chats with that person. A week later, the three were called to simultaneous 2 p.m. meetings and fired. "I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels," said Tomek Korbak, one of the three. The researchers had denied any part in a leak to The Information about OpenAI's model designs, as reported by TechCrunch in AIPD's [October 9th edition](https://aipolicydaily.org/archive/daily/2026-10-09/). [WSJ](https://www.wsj.com/tech/ai/the-signal-chats-that-cost-openai-researchers-their-jobs-594c5853) - **Banks draft recovery plans for large-scale AI attack** — America's largest banks are preparing for the "day after" an AI-driven catastrophe while also working to prevent one, Axios reported. The plans cover protecting accounts, reassuring customers and keeping operations running through a large-scale attack, and scenario planning is already required of critical infrastructure providers, including in the financial sector. Major banks took part in early testing of Anthropic's Mythos, the company's most cyber-capable model, in a joint effort with AI firms to understand and prepare for cyberattacks. JPMorgan CEO Jamie Dimon, speaking in Chicago earlier this month, warned that AI agents may not act properly on a user's behalf. "You can't stop it. You just have to, like, manage the hell out of it," Dimon said. [Axios](https://www.axios.com/2026/10/23/ai-catastrophe-wall-street) - **Microsoft says China-aligned hackers used AI agents to breach 2,250 ConnectWise customers** — Hackers aligned with the Chinese state broke into ConnectWise's cloud backup service and stole credentials that link the IT management company to its corporate customers, Microsoft researchers said, per TechCrunch. With those credentials, the hackers could sign in to customers' Microsoft hosted environments as authorized users and read email and other files. Microsoft said the group used AI agents to break into about 2,250 organizations at a pace of almost 3,000 attempted sign-ins per minute, leaving defenders little time to respond. It attributed the speed to ARTEX, an open source AI framework built to stress-test cyber defenses. ConnectWise paused access to its cloud backup system and said it was working to restore service. [TechCrunch](https://techcrunch.com/2026/10/22/chinese-hackers-spied-on-thousands-of-companies-after-breaching-it-giant-connectwise/) # China Watch - **DeepSeek, MiniMax clear Beijing's new AI export filing and press ahead with overseas licensing** — DeepSeek and MiniMax have completed the registration step under China's proposed AI export rules, Pandaily reported, citing Yuyuan Tantian, a social media account run by state broadcaster China Media Group. The Commerce and Science and Technology ministries opened a draft revision of China's restricted export technology list for comment on Oct. 16, adding 23 AI items covering large models, AI agents and computing clusters. Bloomberg described the step as approval for global use. DeepSeek open-sourced software for running its models on Oct. 20. MiniMax requires organizations in the U.S., EU, UK and South Korea to pass a compliance review and obtain a license before deploying its open models locally. [Bloomberg](https://www.bloomberg.com/news/articles/2026-10-23/deepseek-minimax-win-approval-for-global-use-state-media-says) [Pandaily](https://pandaily.com/deepseek-minimax-complete-ai-technology-export-registration-open-source-overseas-licensing) - **iFlytek releases flagship model it says was trained entirely on Chinese chips** — iFlytek released its Spark X2.6 model family at its annual developer festival, saying the cloud model was trained entirely on domestic computing hardware, Pandaily reported. Chairman Liu Qingfeng said the model was trained over the past two to three months. The company said its method works around the slower memory of Chinese AI chips, cutting one long-document training run from 4.5 months to 1.2 months. iFlytek claims the model ranked first on two software engineering benchmarks, scores it reported itself. Introductory pricing starts at 0.9 yuan (about 13 cents) per million input tokens, the units of text a model processes, and 3.6 yuan (about 51 cents) per million output tokens. iFlytek also open-sourced a small version of the model built to run on robots, cars and PCs. [Pandaily](https://pandaily.com/iflytek-spark-x2-6-domestic-compute-training-open-source-4b-on-device-sparkaudio) - **Unitree founder says company has built 10,000 to 20,000 of one humanoid model since 2024** — Unitree Robotics founder Wang Xingxing said at the company's World Physical AI Conference that physical AI, which uses AI models to control robots and other machines, is showing early signs of a breakthrough, per 36Kr. He cited the release of OpenAI's GPT-6 and said the shift could come within one to three years. Wang named the alignment of AI model inputs and outputs with real robot hardware as the core obstacle to a "ChatGPT moment" for physical AI. He expects robots to reach commercial service settings faster than homes or factories. Wang said Unitree has produced about 10,000 to 20,000 of its G1 humanoid robots since the model launched in May 2024. [36Kr](https://www.36kr.com/p/4037687068021637) # Policy Tracker - **More than 150 midterm ads target data centers and the AI industry** — More than 150 campaign commercials in this year's midterm elections focus on data centers or the AI industry, according to The Guardian. One, paid for by the Democrat-aligned Three Star PAC, accuses Brent Taylor, a Tennessee Republican congressional candidate, of writing the state law that opened Tennessee to data centers. Taylor's campaign answered with its own ad portraying him as tough on tech and data centers. Ken Paxton, a Republican running for U.S. Senate in Texas, has urged a slowdown, breaking with President Trump's pro-AI line. [The Guardian](https://www.theguardian.com/us-news/ng-interactive/2026/oct/22/us-midterms-tech-datacenters-ai) - **OpenAI, Anthropic and Nvidia each post record lobbying quarter** — Nvidia, Anthropic and OpenAI each set a quarterly lobbying record from July through September, spending $2.2 million, $2.1 million and $1.5 million respectively, federal disclosures reviewed by The Verge show. Established players still spend more, with Microsoft at $2 million or more every quarter since 2021 and Google regularly above $3 million. Anthropic's filing names the Kids Online Safety Act's AI sections, the AI Overwatch Act's export control language and the Ratepayer Protection Act, which would make developers bear data center energy costs. It also lists President Trump's executive order renaming AI "Super Intelligence." [The Verge](https://www.theverge.com/policy/1016585/openai-anthropic-nvidia-lobbying-records) - **USDA IT official says staffing cuts rule out human review of every AI task** — Rudolf Rojas, the Agriculture Department's IT program manager and security lead, said Wednesday that cutbacks leave the agency without the staff or contractor support to keep a human in the loop on AI work, per FedScoop. "We just don't have the resources anymore," Rojas said at FedInsider's Innovation Summit in Washington. He said he would let AI agents run without a person only with minimal privileges and a strong risk management framework. Rojas also called for national rules and said he partly rejects the administration's stance, as he characterized it, that AI companies should be free to develop without limits. Ajay Phogat, a senior adviser to the chief information officer at U.S. Customs and Border Protection, backed keeping humans in the loop. [FedScoop](https://fedscoop.com/ai-human-in-the-loop-agency-workforces/) - **Hospital and physician groups tell FDA that developers should answer for generative AI device safety** — The American Hospital Association, American Medical Association and American Telemedicine Association told the Food and Drug Administration that monitoring medical devices built on generative AI would overload providers, especially small and rural practices, Healthcare Dive reported. The groups filed comments on the FDA's proposed approach to regulating such devices, which include chatbots and tools that summarize medical information. They said developers should carry the monitoring burden and the agency should create channels for providers to report problems. Hospitals also told the FDA that limited insight into how the models work makes flaws hard for providers to catch. [Healthcare Dive](https://www.healthcaredive.com/news/developers-not-doctors-held-accountable-generative-ai-medical-devices-fda/833738/) - **Judge weighs Anthropic's fair use defense in music publishers' lyrics suit** — Anthropic asked U.S. District Judge Eumi Lee in San Jose on Wednesday to rule that training its Claude chatbot on copyrighted song lyrics is fair use, per Reuters. Universal Music Group, Concord and ABKCO sued in 2023, and both sides are seeking summary judgment, a ruling without a full trial. Anthropic's lawyer, Sonal Mehta, said the training produces transformative new content and does not reduce demand for lyrics. Publishers' lawyer Matt Oppenheim said users asked Claude to generate lyrics nearly 120,000 times over six months and that models like it "could overwhelm the market with AI-generated content very easily." Lee questioned both sides and did not rule. [Reuters](https://www.reuters.com/legal/legalindustry/anthropic-music-labels-make-key-arguments-ai-training-case-2026-10-21/) [Law360](https://www.law360.com/commercialcontracts/articles/2540717) - **ICE monitoring contract draws scrutiny of DHS AI chief's power to waive surveillance limits** — Immigration and Customs Enforcement (ICE) awarded Thomson Reuters Special Services a $125 million contract in September for services that would enable monitoring of 1 million people, FedScoop reported. The Homeland Security Department's updated AI directive, in effect since September, keeps a ban on improper large scale surveillance or tracking of individuals. But it lets the department's chief AI officer waive any of its requirements, including prohibited uses. An ICE spokesperson said any AI use under the contract will be assessed for consistency with the directive. "A rule that comes with its own off switch is not a rule," said Tom Bowman, policy counsel at the Center for Democracy & Technology. [FedScoop](https://fedscoop.com/dhs-ai-policy-updates/) # Capability & Research Watch - **Hackers hijack LiteLLM release pipeline to push credential stealing malware** — Attackers compromised the release process of LiteLLM, an open source tool that lets apps reach AI models from many providers through one interface, and published a malicious version 1.107.0 to PyPI, the main Python software index, and Docker on Wednesday, BleepingComputer reported. The project logs 92 million monthly downloads from PyPI and is used by companies including Nvidia, Netflix and Okta. The malware steals cloud credentials, AI provider keys, developer secrets and browser data, and runs every time Python starts in an affected environment. Security firms flagged the release within six minutes of publication. The maintainer said customers running its private enterprise version were not affected. [BleepingComputer](https://www.bleepingcomputer.com/news/security/litellm-release-pipeline-hacked-to-distribute-info-stealer-malware/) - **Microsoft says its AI diagnosis system beat generalist doctors on complex cases** — Microsoft's MAI-DxO, which coordinates several AI agents acting as a panel of doctors, reached the right diagnosis in 83% of complex New England Journal of Medicine cases in a simulation, against 36% for primary care doctors and general internists working alone, TechCrunch reported on a study published in Science. Specialists, who usually handle such cases, were not part of the comparison. Much of the gain came from stronger underlying models, and Microsoft's best results used OpenAI's GPT-5 and GPT-5.5. On emergency department cases, the system added little accuracy or cost savings. Microsoft AI CEO Mustafa Suleyman called the result a step toward "medical superintelligence." [TechCrunch](https://techcrunch.com/2026/10/22/microsofts-mustafa-suleyman-says-its-ai-outperforms-doctors-on-hard-cases-as-he-chases-medical-superintelligence/) # Industry & Market Watch - **Major banks restrict AI agents after spotting unauthorized ones in customer accounts** — Large U.S. banks are finding third-party AI agents operating inside customer accounts without authorization. Some agents appear to act independently of user instructions, people familiar with the matter told Axios. Morgan Stanley has begun blocking outside agents from its platforms, and JPMorgan, Bank of America and Citi are also restricting them. Many banks cannot tell whether agent activity in an account was directed by the customer. Some are planning how to quickly secure or restore access if an agent goes rogue. A Morgan Stanley spokesperson said the bank has not seen rogue AI agents on its platforms. Most banks allow, or are moving toward, read-only access that lets agents view account data without acting on it. [Axios](https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access) - **OpenAI close to raising up to $30 billion from UAE-led group at $1.25 trillion valuation** — OpenAI is close to raising as much as $30 billion at a $1.25 trillion valuation before the new money, in a round led by investors from the United Arab Emirates and Coatue, CNBC reported. That is below the $1.4 trillion figure Bloomberg reported a month ago, and terms are still being negotiated. OpenAI told investors it aims to collect signatures by the end of next week. It is also negotiating a deal for more computing capacity from the UAE. OpenAI is positioning the round as a bridge to a potential initial public offering, which Chief Financial Officer Sarah Friar has told employees could happen in 2027. OpenAI's March financing of $122 billion valued it at $852 billion. [TechCrunch](https://techcrunch.com/2026/10/22/openai-working-on-new-30b-raise-at-1-25t-valuation-ahead-of-ipo/) [CNBC](https://www.cnbc.com/2026/10/22/openai-funding-round-uae-investors.html) - **IDC projects FCC foreign robot rule will cut U.S. humanoid revenue 58% by 2030** — Research firm IDC projects that U.S. humanoid robot revenue will fall 58% below its expected level by 2030, Forbes reported. IDC blames the Federal Communications Commission's July 28 addition of new foreign-made advanced robots to its Covered List, which blocks the authorization they need for U.S. sale. Humanoids take the biggest hit because their makers release new models quickly and lack a catalog of older certified ones, IDC said. The firm expects the UK, France, Germany and Eastern Europe to absorb most of the redirected robots, and sees more than $6 billion in lost U.S. robotics revenue from 2026 to 2030. "Most coverage calls this a ban, but in practice it is a chokepoint, not a wall," IDC analysts Navkendar Singh and Ryan Reith wrote. [Forbes](https://www.forbes.com/sites/johnkoetsier/2026/10/22/fcc-robot-ban-cuts-us-humanoid-robot-revenue-58-send-chinese-robots-to-europe-idc/) - **Goldman economist projects AI could displace up to 9% of workers, mostly during downturns** — Goldman Sachs economist Joseph Briggs estimated in a note that up to 9% of workers could lose their jobs during the AI transition, with slower hiring cutting employment by another 3%, Fortune reported. The forecast assumes AI eventually lifts productivity by 15% in the U.S. and other developed economies. Briggs found that companies have shed workers to automation mainly in past recessions, while the share of Americans in routine jobs held steady during expansions. He concluded that the timing and extent of AI driven job losses will likely track the broader economic cycle. [Fortune](https://fortune.com/2026/10/22/goldman-sachs-forecast-how-many-jobs-lost-to-ai-9-percent/) # Global & Geopolitics - **UK pulled 32 government services offline over PHP flaws, minister tells Parliament** — Britain's cybersecurity minister, Liz Lloyd, told Parliament that 32 government services had been taken offline over the PHP flaws, mostly as a precaution while updates were installed, per The Guardian. Downing Street said the outages had not reached any critical public facing services. The National Cyber Security Centre, Britain's cyber defense agency, sent 50,000 alerts to affected organizations and said it had no reports of successful hacks. It also said the patch will not work on some older systems, which will have to be replaced. [The Guardian](https://www.theguardian.com/technology/2026/oct/22/websites-across-world-hit-by-millions-of-attempts-to-exploit-php-code-flaw) - **Fewer than a quarter of planned Sydney and Melbourne data centers have confirmed power** — Of the more than 40 large data centers planned or under construction around Sydney and Melbourne, representing more than A$100 billion ($70 billion) in projected investment, just 23% have publicly confirmed electricity supply or grid connections, per a Bloomberg review of company disclosures. The shortfall puts at risk Australia's aim of becoming an AI hub for the Asia-Pacific region. In Sydney, seven of 28 active projects have confirmed power, against three of 15 in Melbourne. Bloomberg contacted 25 developers, and apart from two operators that gave statements, nine declined to comment or did not answer directly. [Bloomberg](https://www.bloomberg.com/news/articles/2026-10-22/ai-data-center-backlash-threatens-australia-s-70-billion-dream) - **UK to set up AI unit in its crisis center to spot wildfires** — Prime Minister Andy Burnham will announce a wildfire AI unit in the National Situation Centre, part of the government's Cobra emergency briefing rooms, at the government's first wildfire resilience summit, The Guardian reported. The government said the unit would help it detect and fight fires faster. Google, Airbus and VodafoneThree are expected to attend and show how satellite monitoring and early warning systems can help. Burnham will also announce £45 million ($60 million) for fire crews to buy wildfire-specific protective gear. [The Guardian](https://www.theguardian.com/world/2026/oct/22/burnham-plans-ai-unit-spot-fight-fires-faster-record-breaking-summer-heat)