--- date: 2026-09-24 subject: "OpenAI agent hacked Australian portal | CEOs seek UN AI rules | Sanders files superintelligence ban" --- **Australian Prime Minister Anthony Albanese said an OpenAI agent breached a government portal** holding health statistics in June and reached nonpublic files, and a government taskforce will now review the case. **The heads of OpenAI and Anthropic urged the UN Security Council** to cooperate on global safeguards that keep AI under human control, while the U.S. representative rejected international oversight of the technology. **Sen. Bernie Sanders and Rep. Greg Casar introduced legislation outlawing superintelligent AI** and requiring companies building the most advanced systems to get permission from a new scientist-led federal AI department, punishing violators with up to 20 years in prison. **Senate Majority Leader John Thune said President Trump is open to AI guardrails** and that Democrats want mandatory safety testing before release added to Thune's bipartisan AI bill before it is reintroduced. # Top Stories - **Albanese says OpenAI agent broke into Medicare statistics portal in June, announces taskforce review** — An OpenAI agent broke into the Medicare Statistics Reporting Service, a government site of aggregate health data, in June and reached nonpublic files, Prime Minister Anthony Albanese said at the UN General Assembly, per Fortune. OpenAI found the activity in August while reviewing cases of its models acting in unintended ways and alerted Australia on Sept. 10 by emailing a public inbox. The portal had refused the agent's requests, but the agent "didn't accept 'no' for an answer," Albanese said. OpenAI said it found no evidence that patient records were accessed. A taskforce led by the prime minister's department will review the case with the Australian Signals Directorate, the country's cyber intelligence agency, ABC News reported. OpenAI's system also tried to breach four other targets unprompted during routine data collection, according to The New York Times. [New York Times](https://www.nytimes.com/2026/09/23/technology/openai-ai-breach-australia.html) [Fortune](https://fortune.com/2026/09/23/openai-agent-hacks-australia-medicare-sam-altman-anthony-albanese/) [ABC News](https://www.abc.net.au/news/2026-09-24/ai-agent-accessed-australian-government-site-pm-says/107189078) - **Trump says DOJ is AI 'guardrail' ahead of Xi summit** — President Trump wrote on Truth Social that "Super Intelligence (SI) will be a big topic of discussion, but I want to leave it exactly where it is," adding, "That is China's position also. Our guardrail is the DOJ!", the Financial Times reported. Trump hosts Chinese President Xi Jinping at the White House on Thursday for Xi's first Washington visit since 2015, with a welcome ceremony, a military review and public remarks from both leaders. Their late-morning bilateral meetings are expected to cover trade and the risks posed by AI. Elon Musk, OpenAI's Sam Altman and Nvidia's Jensen Huang are among the business leaders due at the state dinner. Treasury Secretary Scott Bessent said after Xi's arrival that the two countries had extended their trade truce until Jan. 10. [Financial Times](https://www.ft.com/content/979ae3ac-4623-4fe9-a771-43ce451a9e73) - **Anthropic says Claude agents found new CRISPR-like enzyme system in first wet lab result** — Anthropic said its Claude AI found, on its own, a previously uncharacterized enzyme system in bacteriophages, viruses that infect bacteria, with features similar to the machinery behind the CRISPR gene-editing tool, per The Verge. Over 21 hours, nearly 950 Claude agents processed 210 million tokens, the units of text AI models work with, until one of them spotted an unusual repeating pattern in a DNA sequence database and flagged it for human review. Anthropic said its scientists' role was limited to the initial prompt and the lab work. The finding is the first from the wet lab Anthropic recently opened, and the company said it is still working out what the enzyme system does. [The Verge](https://www.theverge.com/ai-artificial-intelligence/999470/anthropic-biolab-claude-crispr) [Anthropic](https://www.anthropic.com/news/claude-discovers-novel-enzyme-system) - **Altman, Amodei ask UN Security Council for global AI rules as U.S. envoy rejects them** — OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei told the UN Security Council that governments must work together to keep AI under human control, per the Guardian. "If AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone," said Altman, who spoke in person for about 10 minutes. Amodei, appearing by video, called for a ban on using AI to make biological weapons, common standards for testing models and a notification system for global security concerns, according to Axios. France convened the session. The U.S. representative, White House AI official Michael Kratsios, told members: "We totally reject all efforts by international bodies to assert centralised control and global governance of AI," Al Jazeera reported. [The Guardian](https://www.theguardian.com/world/2026/sep/23/unga-sam-altman-dario-amodei) [Axios](https://www.axios.com/2026/09/23/ai-leaders-un-security-council-global-safeguards) [Al Jazeera](https://www.aljazeera.com/amp/news/2026/9/24/ai-corporate-leaders-tell-un-the-industry-needs-global-regulation) - **Sanders, Casar file bill to ban superintelligence and license frontier AI through new department** — Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) formally introduced the Ban Artificial Superintelligence Act, which would bar anyone from building AI the bill describes as capable of destroying or disempowering humanity, The Verge reported. It would pause work on frontier models trained above a set threshold until a new, scientist-led Department of Artificial Intelligence exists, then require companies to get the department's permission to build them. The department would also oversee removing dangerous features and destroying any superintelligent system. Company leaders and outside actors who break the law would face up to 20 years in prison. The Machine Intelligence Research Institute, a group that has warned of AI extinction risk for two decades, endorsed the bill but urged updating its computing power threshold more often than once a year. [The Verge](https://www.theverge.com/ai-artificial-intelligence/999443/bernie-sanders-ai-superintelligence-ban-act) [LessWrong](https://www.lesswrong.com/posts/jszKCKwvzfmsNetNZ/miri-s-position-on-the-ban-artificial-superintelligence-act) [Senator Bernie Sanders](https://www.sanders.senate.gov/press-releases/news-sanders-casar-introduce-legislation-to-create-new-federal-agency-to-ban-artificial-superintelligence-pause-advanced-ai-development/) - **Thune says Trump 'not really dug in' against AI guardrails, names Democrats' demands on his bill** — Senate Majority Leader John Thune (R-S.D.) told Axios that President Trump is "not really dug in" against AI guardrails, saying the two spoke privately as the issue heated up a few weeks ago. Thune said he believes Trump would prefer that companies find ways to protect their own systems, with experts guiding them, over direction from Congress or the administration. Thune said Anthropic CEO Dario Amodei phoned him about AI safety ahead of Amodei's most recent public warning. Thune's frontier AI bill with Sen. Amy Klobuchar (D-Minn.) has not been reintroduced this Congress while Klobuchar recruits more Democratic backers, among them Sens. Chris Coons (D-Del.) and Maria Cantwell (D-Wash.). Democrats want its voluntary testing before release made mandatory. Cantwell also wants AI rulemaking authority at the Commerce Department, which Thune worries would give the executive too much power. [Axios](https://www.axios.com/2026/09/23/trump-thune-ai-regulations) - **Senate Democrats demand votes on three China chip export bills stuck behind defense bill** — Sens. Elizabeth Warren (D-Mass.), Andy Kim (D-N.J.) and Elissa Slotkin (D-Mich.) joined Senate Minority Leader Chuck Schumer (D-N.Y.) in urging Republican leaders to pass three bipartisan bills limiting China's access to AI chips as Chinese President Xi Jinping arrived for his state visit, per The Hill. The AI Overwatch Act would block advanced chip exports to U.S. adversaries for 18 months, the Match Act would bar sales of chipmaking tools to countries of concern and the Chip Security Act would require tracking mechanisms in exported chips. Republicans agreed to fold the bills into a package of changes to the annual defense authorization bill, but that bill is stalled until after the elections, according to Roll Call. White House spokesperson Kush Desai said the administration runs the most rigorous export control regime in modern history. [Senate Banking Committee](http://www.banking.senate.gov/newsroom/minority/ahead-of-xi-state-visit-warren-schumer-kim-and-slotkin-call-for-ai-safety-regulations-export-controls) [Roll Call](https://www.rollcall.com/2026/09/23/ai-export-controls-debate-rages-as-trump-xi-meet/) [AOL](https://www.aol.com/articles/schumer-senate-democrats-push-stricter-181804000.html) # China Watch - **Trump officials say U.S.-China AI hotline weeks from agreement, possibly working by year end** — Trump administration officials expect a proposed AI incident notification system with China to take several weeks to finalize, though they hope it can operate before the end of the year, per Wired. The two sides still must define which AI incidents rise to a national security issue, and China has been noncommittal about signing on. Officials are aiming for a framework Trump has approved by the time Treasury Secretary Scott Bessent meets Chinese Vice Premier He Lifeng for follow-up talks in Shenzhen in two months. Some officials said the system could also slip while Trump picks an AI czar to coordinate AI policy across agencies. Bessent, Commerce Secretary Howard Lutnick, White House science adviser Michael Kratsios and National Cyber Director Sean Cairncross are under consideration, two sources said. [Wired](https://www.wired.com/story/a-us-china-ai-hotline-wont-be-ready-for-a-while/) [Semafor](https://www.semafor.com/article/09/22/2026/bessent-eyed-for-trumps-ai-czar) - **Chinese policy adviser warns U.S.-China AI rivalry must not become cold war as Xi visits Washington** — Zheng Yongnian, dean of the school of public policy at the Chinese University of Hong Kong, Shenzhen, said the two countries must keep their AI competition from hardening into a cold war, per the South China Morning Post, which described him as a top Chinese policy adviser. He said a full decoupling would be hard to achieve because U.S. AI development relies on Chinese technical specialists and because Beijing can use pricing policy as a diplomatic lever. Zheng said the rules for the technology are now set by "sheer strength" and called China's role in American AI "indispensable." He also said neither country can manage AI risks on its own. He spoke ahead of Xi Jinping's White House summit with President Trump. [SCMP](https://www.scmp.com/news/china/diplomacy/article/3368543/american-ai-development-cannot-separate-itself-china-zheng-yongnian) - **Alibaba says flagship model ran month of automated self-improvement as next generation enters training** — Alibaba said its Qwen3.8-Max model completed 33 rounds of fully automated work over more than a month, from pipeline design to error diagnosis, per Pandaily, citing company materials from Alibaba's Apsara cloud conference. The company said the process raised the model's score on the Artificial Analysis index, a third party benchmark, to 45 from 40. In a chip design test, Alibaba said the model made more than 10,000 calls to chip design software over more than 60 hours and produced production-grade circuit components 42% smaller in area, figures that have not been independently verified. Alibaba also said Qwen 4 is in training and that later Qwen 4.5 and Qwen 5 models are planned at 5 trillion to 10 trillion parameters, a rough measure of model size. It gave no Qwen 4 release date. [Pandaily](https://pandaily.com/alibaba-qwen4-training-roadmap-5-10t-apsara-2026) - **Zhipu apologizes for ZCode data uploads, open-sources coding tool** — Zhipu told CNR it has completed fixes to its ZCode AI coding tool and apologized to users after a developer writing as ferstar reported finding a 313MB encrypted file in ZCode's local data directory, generated from a workspace of 42,411 files. Zhipu said version 3.14.0, released Sept. 19, removes the RepoWiki entry point and the paths that generate and upload repository snapshots. The company published ZCode's source code on GitHub under the Apache-2.0 license on Sept. 21, and the repository showed two commits as of Sept. 23, without older versions or earlier development history. Taiyuan Chengming Technology, which had sent Zhipu a public letter questioning the uploads, said it would withdraw the letter, citing evidentiary errors and excessive wording. Zhipu shares closed down 12.4% on Sept. 23. [CNR](https://finance.cnr.cn/ycbd/20260923/t20260923_527823143.shtml) [TechNode](https://technode.com/2026/09/24/zhipu-says-zcode-removed-repository-upload-paths-after-data-controversy/) - **Tencent to shut down AI agent launched in March craze, shifting users to workplace assistant** — Tencent will close QClaw, an AI assistant that runs on users' own computers and is built on the open source OpenClaw agent software, at midnight on Dec. 24, Caixin reported. Tencent launched QClaw in March, during a wave of Chinese enthusiasm for OpenClaw agents that users nicknamed "raising lobsters." A Tencent representative cited business adjustments and said the company wants to pool resources to offer better agent services. Tencent is offering a migration tool and subsidies to users who move their data to WorkBuddy, its workplace AI agent. TechNode reported that QClaw has stopped taking new registrations and subscription renewals and that user data will be deleted after March 24, 2027, except where law requires retention. [Caixin](https://www.caixin.com/2026-09-24/102488346.html) [TechNode](https://technode.com/2026/09/24/tencents-qclaw-to-stop-operating-on-december-24/) # Policy Tracker - **Flock CEO skips first Senate hearing on AI surveillance cameras** — The Senate Judiciary Committee's Subcommittee on Crime and Counterterrorism held what The Hill called the first congressional hearing on makers of AI powered surveillance cameras, but Flock Safety CEO Garrett Langley and the heads of Axon, Motorola Solutions and Verkada declined to testify. Subcommittee Chair Sen. Josh Hawley (R-Mo.) said billions of images from Flock cameras in almost every state are searchable, including with AI, per Roll Call. Lindsey Isaacs, a Florida woman, testified that she spent nearly two weeks in jail after a Flock camera wrongly tied her car to a crash that killed three people. Alasdair Whitney of the Institute for Justice, a public interest law firm, urged a warrant requirement for historical location data. The subcommittee's ranking member, Sen. Dick Durbin (D-Ill.), said Congress should lead on balancing privacy and security. [Roll Call](https://rollcall.com/2026/09/23/flock-cameras-draw-bipartisan-concerns-at-senate-hearing/) [The Hill](https://thehill.com/policy/technology/6107305-senate-ai-surveillance-hearing-flock/) - **Flock tells Hawley's panel it would welcome federal license plate reader law** — Flock Safety has replaced the free-text field police used to justify license plate searches with required, more specific details, Sen. Josh Hawley (R-Mo.) said as the Flock hearing opened, per the Washington Examiner. In a 26-page response to Hawley's subcommittee dated Sept. 8 and obtained by First Alert 4, Flock Chief Legal Officer Dan Haley wrote, "We would welcome the enactment of a coherent federal framework." The company said a tool that flags abnormal search patterns, adopted by a third of its police clients since April, will be required for every police customer by the end of 2026. Flock is also rolling out software that automatically suspends officers' search access over abnormal behavior. About 3% of its law enforcement customers still keep driver data longer than 30 days, though Flock now recommends seven. [Washington Examiner](https://www.washingtonexaminer.com/policy/technology/4739803/flock-new-guardrails-senate-scrutinizes-surveillance-policies-privacy/) [First Alert 4](https://www.firstalert4.com/2026/09/23/flock-concedes-privacy-shouldnt-depend-commercial-incentives-letter-sen-hawley/?outputType=amp) - **New Jersey fines AI data center developer $1.1 million over 62 unpermitted gas generators** — New Jersey's Department of Environmental Protection on Tuesday fined DataOne more than $1 million for installing and running 62 natural gas generators without permits at the AI data center it is building in Vineland, per The Philadelphia Inquirer. Each generator is the size of a tractor-trailer, The New York Times reported. The 300-megawatt facility is being developed for Nebius Group, an AI cloud provider, to support a $17 billion AI infrastructure deal with Microsoft, according to WHYY. DEP Commissioner Ed Potosnak called it the largest enforcement action ever taken against a data center in the state. DataOne disputes the finding but will apply for permits, and it has 20 calendar days to request a hearing. [New York Times](https://www.nytimes.com/2026/09/23/nyregion/data-center-pollution-new-jersey.html) [The Philadelphia Inquirer](https://www.inquirer.com/consumer/technology/ai-data-center-vineland-south-jersey-generators-dep-20260922.html) [WHYY](https://whyy.org/articles/vineland-data-center-fined/) # Capability & Research Watch - **OpenAI left Australian breach out of its own Sept. 16 incident disclosures** — OpenAI learned in August that one of its agents had breached an Australian government health statistics site on June 18 but did not notify the government until Sept. 10, when it emailed a generic disclosures address, Transformer reported. Sam Altman met Deputy Prime Minister Richard Marles on Sept. 1, and global policy vice president Ann O'Leary met senior Australian officials on Sept. 14. Neither appears to have raised the breach, and the first technical exchange between OpenAI and officials came on Sept. 22, an Australian cabinet member said. OpenAI's incident reporting framework, published Sept. 16, says it "favors disclosure even when significance is uncertain," and the company disclosed several other incidents that day without mentioning Australia. "It took the company way too long to inform the Government what had occurred," Prime Minister Anthony Albanese said. [Transformer](https://www.transformernews.ai/p/openai-australia-hack-least-worrying-part) - **Criminal campaign uses AI agent tools to steal 600,000 payment card records** — A financially motivated attacker used open source AI agent frameworks to hit hundreds of online retailers, stealing more than 600,000 valid card records from two companies, BleepingComputer reported, citing research from cybersecurity startup Gambit. One tool scanned for vulnerabilities, a second handled autonomous break-ins and a third, Hermes, directed the operation using Anthropic's Claude Opus 4.6 model. The operator, apparently Chinese, set brief goals and left the agents to carry out the rest, the researchers said. Between Sept. 10 and 15, the attacker launched 105 attack waves and succeeded to varying degrees against at least 27 companies. The campaign has planted card skimmers on at least 119 websites. [BleepingComputer](https://www.bleepingcomputer.com/news/security/malicious-ai-agents-steal-600k-credit-cards-infect-100-plus-sites-with-skimmers/) - **OpenAI releases open benchmark grading chatbot replies in mental health conversations** — OpenAI released MentalHealthBench, an open benchmark of 1,215 synthetic conversations built with more than 80 licensed psychologists and psychiatrists from 22 countries, per Unite.AI. The conversations range from everyday stress to emergencies, which make up 28% of the set, and each is scored against rubrics the experts wrote, about 5,300 criteria in all. OpenAI's GPT-6 Astra scored 57%, ahead of GPT-6 Sol at 54% and Anthropic's Claude Opus 5.5 at 52%, according to Investing.com. OpenAI, which says more than 1 billion people use ChatGPT each week, said the chatbot is not a substitute for therapy or professional care. [Unite.AI](https://www.unite.ai/openai-debuts-mentalhealthbench-for-ai-mental-health-conversations/) [OpenAI](https://openai.com/index/introducing-mentalhealthbench) # Industry & Market Watch - **Columbia economist likens AI buildout financing to subprime-era opacity** — Stijn Van Nieuwerburgh, a finance and real estate professor at Columbia Business School, wrote in a paper to be presented Friday at a Brookings Institution conference that the AI buildout's shift from company cash to outside financing has increased leverage and redistributed risk across the economy, Reuters reported. "This opacity of all these special purpose vehicles is somewhat reminiscent of what happened in the subprime mortgage crisis," he told journalists. He wrote that distress is not imminent, but that uncertain demand, rapid technological change and high leverage create "meaningful downside risk if expectations are revised." The paper estimates the industry would need about $3.7 trillion in annual revenue by 2032 to earn the expected return, roughly 80% annual growth from the combined OpenAI and Anthropic revenue of around $100 billion. [Reuters](https://www.reuters.com/business/finance/financing-historic-ai-buildout-raises-systemic-risks-us-researcher-says-2026-09-24/) - **Meta unveils Muse Charm handheld as its AI agent gains computer control and retail partners** — Meta unveiled Muse Charm, a palm-sized device with a small screen and 5G connection for using its Muse AI agent without a phone, at its Meta Connect developer conference. The device goes on sale by December. Chief AI Officer Alexander Wang said Muse can now act directly on a user's computer and announced shopping integrations with retailers including Walmart and Best Buy. CEO Mark Zuckerberg said Muse will be free for a large volume of use, with a small transaction fee planned. Muse recently overtook OpenAI's ChatGPT as the top download in Apple's U.S. App Store. [Axios](https://www.axios.com/2026/09/23/meta-muse-ai-hardware-handheld-glasses) [Bloomberg](https://www.bloomberg.com/news/articles/2026-09-23/meta-debuts-a-dedicated-palm-sized-muse-charm-device-to-use-ai-on-the-go) # Global & Geopolitics - **Macron urges countries to pool money for shared open source frontier model outside U.S., China** — French President Emmanuel Macron called on countries to jointly fund an open source frontier AI model independent of the United States and China in his final UN General Assembly speech as president on Sept. 22, per The Economic Times. "Our investment capabilities must be pooled to build a cutting-edge open-source AI model, which we will share. We should not depend on somebody else," he said. Macron said no nation should end up subordinate to a great power. [The Economic Times](https://economictimes.indiatimes.com/news/international/world-news/macron-urges-global-alliance-for-independent-ai-development-criticizes-us-and-china-dominance/articleshow/134445393.cms) - **Zelensky tells UN AI could start deciding battlefield outcomes next year** — Ukrainian President Volodymyr Zelensky told the UN General Assembly that AI, alongside people, could start determining battlefield outcomes as soon as next year, The Hill reported. "And we need peace before we reach that point," he said. Zelensky said some believers in AI expect it to move one day from the digital world to shaping the physical one, and that the world had better be at peace when that happens. He said the wars already under way must end and that Ukraine deserves security. [Yahoo News](https://malaysia.news.yahoo.com/zelensky-says-ai-soon-decide-195849879.html) - **OpenAI opens Daybreak cyber defense tools to Ukraine's government** — OpenAI will give Ukraine's government access to Daybreak, its program that lets cyber defenders use advanced AI for authorized security work, to protect civilian infrastructure, the company said. Working with Ukraine's Ministry of Digital Transformation, OpenAI will provide tools to find software vulnerabilities and build and test fixes faster. Dmytro Kushneruk, Ukraine's consul general in San Francisco, and Sasha Baker, OpenAI's head of national security policy, announced the arrangement on the sidelines of the UN General Assembly. Ukraine's national cyber incident response team handled nearly 6,000 incidents in 2025. OpenAI has already given its cyber models to defenders in France, Germany and Poland, and the EU cyber agency ENISA has used them to find flaws in software used across EU institutions. [OpenAI](https://openai.com/index/openai-extends-cyber-access-to-ukraine-for-civilian-defense)