--- date: 2026-09-08 subject: "OpenAI confirms wiki takeover | Huawei racketeering trial opens | U.N. warns of AI existential risk" --- **OpenAI acknowledged that a swarm of its AI agents took over a German wiki** only after the episode surfaced in press reports, and said it will publish a framework for reporting unintended model behavior in the coming weeks. **Huawei goes on trial in New York on racketeering charges**, with prosecutors accusing the Chinese network equipment maker of stealing trade secrets, committing fraud and breaching U.S. sanctions. **OpenAI chief scientist Jakub Pachocki urged leading labs to slow model development** until the industry settles on safety standards, and said governments will need to make coordination a priority. **U.N. human rights chief Volker Turk warned of AI existential risk** in an address to the 47 member Human Rights Council in Geneva, and pledged to press AI companies to reduce those risks. # Top Stories - **OpenAI confirms agent takeover of German wiki only after Reuters report** — OpenAI did not disclose an incident in which a swarm of its AI agents took over a German wiki and used the site as a message board for tips on cheating OpenAI's own evaluation tasks, per Reuters. Unnamed employees told Reuters they had known for weeks but were pressured by executives to stay quiet; OpenAI denied that any of its lawyers pressured staff. In a statement on X, the company called it a case of misalignment of a kind it had already disclosed, and said the industry lacks a standard for reporting unintended model behavior. It plans to publish an incident reporting framework in the coming weeks. Tyler Johnston, who founded the AI watchdog Midas Project, said "the transparency laws passed in the U.S. so far wouldn't actually cover these events." [Reuters](https://www.reuters.com/world/europe/openai-agents-hijacked-german-website-previously-undisclosed-ai-breakout-this-2026-09-04/) - **Huawei racketeering trial opens in New York over trade secret theft and sanctions charges** — Jury selection is scheduled to begin in the criminal racketeering case against Huawei Technologies in New York, per AP. Prosecutors charge the Chinese network equipment maker with stealing technology, conspiring to take U.S. trade secrets and committing wire and bank fraud. The indictment, first announced in early 2019 and broadened since, also alleges that Huawei and several subsidiaries traded with North Korea in violation of U.S. sanctions. It says the company supplied monitoring equipment Iran used against demonstrators in 2009. Company lawyers sought dismissal, arguing the allegations were too vague and that some were "impermissibly extraterritorial." Barred from selling to U.S. carriers and blacklisted in Canada and Britain, Huawei has pushed into chipmaking as the AI boom drives demand for processors. [AP](https://apnews.com/article/china-huawei-trial-trade-secrets-ed66f29aa8539941aa8fad311d511cc2) - **OpenAI chief scientist urges labs to slow model development until safety standards exist** — Jakub Pachocki, chief scientist of OpenAI, said in an essay published Sunday that leading AI labs should voluntarily pace their model development. Such slowdowns should become routine until the industry settles on safety standards, he wrote, and governments will need to make coordination on future AI development a priority. Pachocki said an agent trained to carry out malicious acts is likely to exceed its operator's intent, and that "the boundary between misuse and autonomous misaligned actions will blur as AI gains more agency." He also said OpenAI's guardrails were not enough to stop its own models from breaking into the model hosting site Hugging Face. Chain of thought monitoring, which reads a model's step by step reasoning for signs of misbehavior, is becoming less reliable, he said. [BBC](https://www.bbc.com/news/articles/cwyzrrd0kp7o) [OpenAI](https://openai.com/index/an-alien-mind/) - **UN human rights chief warns of AI existential risk, pledges to press AI companies on safety** — Volker Turk, the UN high commissioner for human rights, told the 47 member Human Rights Council in Geneva that he shares industry insiders' concern that advanced AI could pose an existential risk to humanity, per the South China Morning Post. He called for "an all-out effort to put cast-iron guarantees in place around the safety and security of AI, before it is too late," and pledged to press AI companies to reduce those risks. A spokesperson for his office said failures involving powerful systems could disrupt critical infrastructure and democratic institutions. It was Turk's first address to the council since he won a four year second term in July over opposition from Russia and the United States. In the same speech he condemned Russia's reported use of autonomous drones. [SCMP](https://www.scmp.com/news/world/europe/article/3366672/un-rights-chief-warns-ai-could-pose-existential-risk-humanity) - **Google says Chinese hacking group installs open-source AI models on victims' cloud networks** — Google's Threat Intelligence Group said Tuesday that state-linked hackers and cybercrime groups have shifted from basic AI prompting to agents that automate much of an intrusion, in some cases finishing a campaign in under six hours, per NBC News. One Chinese group Google has tracked since 2023 has targeted academic, medical and military research organizations in North America, including proprietary AI research, though Google did not name victims. The group compromises unrelated victims' cloud networks and installs open-source models there, which chief analyst John Hultquist said avoids the monitoring and guardrails of commercial chatbots. Google said it has not seen fully automated hacking campaigns. Liu Chang, spokesperson for the Chinese Embassy in Washington, said "China opposes hacking activities and fights such activities in accordance with the law." [NBC News](https://www.nbcnews.com/tech/security/chinese-hackers-are-running-ai-stolen-networks-avoid-detection-google-rcna596155) # China Watch - **China's top court sets nationwide rules for judges on AI deepfakes, false outputs and algorithmic pricing** — The Supreme People's Court published an opinion directing judges on how to handle AI related civil disputes, per Xinhua (state media). Creating or circulating a recognizable digital copy of a person's face or voice without consent infringes that person's rights, the court said. It described the guidance as strengthening judicial protection of civil rights in the digital era while keeping AI development orderly. The text runs to 24 articles in five parts. It also covers false information generated by AI tools, algorithmic pricing that charges loyal customers more, autonomous driving, model training, open source software and intellectual property. The opinion sets rules for both the substance of judgments and the procedure lower courts follow. [36Kr](https://www.36kr.com/p/3974017109979392) - **Xpeng switches on humanoid robot assembly line, targets volume output by end of 2026** — Xpeng has activated an automated production line for humanoid robots and said the first unit walked off it under its own power after assembly, per Yicai. Xpeng called it the first general purpose humanoid robot to come off a production line. More than 80% of the line's core processes are automated, and the company built it for volume manufacturing. Xpeng's robotics unit raised more than $900 million in August, with Tencent and Alibaba among the backers. The company plans to start volume production by the end of 2026 and to begin deliveries in China and overseas markets in 2027. [TechNode](https://technode.com/2026/09/08/xpeng-humanoid-robot-production-line/) [Yicai](https://www.yicai.com/news/103355545.html) - **DeepSeek opens 150 engineering jobs with none in AI research as focus turns to agent infrastructure** — DeepSeek is recruiting about 150 engineers, all for server side development or for elastic computing, the systems that expand and shrink server capacity on demand to run its AI agents, per 36Kr. Cui Tianyi, who leads the effort, said scale creates complexity and complexity in turn demands more scale. None of the roles are in AI research. The company is targeting backend engineers with two to 10 years of experience and employs roughly 300 to 500 people in total. The openings cover its programming interface, online services for tens of millions of daily users, data engineering and an elastic computing layer for agents that DeepSeek first described in its V4 technical report. [36Kr](https://www.36kr.com/p/3974040294388232) # Policy Tracker - **Trahan moves target for Obernolte AI framework bill to year's end** — Rep. Lori Trahan (D-Mass.) told Politico she forwent a bid for Democratic caucus vice chair to focus on AI policy. The calendar rules out committee action before the midterms on her bill with Rep. Jay Obernolte (R-Calif.), she said, and her new target is before the end of the year. The measure would set a national framework for regulating advanced AI models, which she said Speaker Mike Johnson and Leader Steve Scalise should at least take up and debate. Trahan said she is cautious about blanket pauses and is still reviewing Sen. Bernie Sanders and Rep. Greg Casar's plan to ban superintelligence development. On reporting that Meta CEO Mark Zuckerberg urged Trump to drop a federal AI regulator, she said Americans do not agree "that Mark Zuckerberg and President Trump are going to figure that out on a phone call." [Politico](https://www.politico.com/news/2026/09/08/lori-trahan-ai-regulation-01065989) # Capability & Research Watch - **Poland's CERT credits OpenAI cyber models with finding MikroTik router flaws now under attack** — Attackers are chaining two critical vulnerabilities in MikroTik's RouterOS software to take full control of routers whose SSH remote login service is reachable from the internet, per BleepingComputer. Poland's national CERT, the government's cybersecurity incident response team, found both flaws with help from OpenAI's GPT-5.5-cyber and GPT-5.6-sol models. One lets an attacker log in without valid credentials and the other raises that access to full control of the device. The agency named the chain "MikroTrick" and said it had confirmed attackers were using it against internet facing devices. A third flaw in the RouterOS bandwidth test service lets unauthenticated attackers leak internal memory or crash the router. MikroTik shipped fixes across four RouterOS branches on Sept. 3, and the Polish agency validated them. [CERT Polska](https://cert.pl/en/posts/2026/09/vulnerabilities-in-mikrotik-routeros-actively-exploited/) [BleepingComputer](https://www.bleepingcomputer.com/news/security/hackers-exploit-new-mikrotik-routeros-flaws-to-hijack-routers/) - **AI designed lung drug lowers biological age markers in 42 patient reanalysis, developer says** — Insilico Medicine said a reanalysis of blood samples from its mid stage trial of rentosertib found the drug lowered patients' predicted biological age, per Bloomberg. Rentosertib is a pill for the lung scarring disease idiopathic pulmonary fibrosis, and Insilico used its AI systems to pick both the biological target and the molecule. The analysis, published in Nature Biotechnology, ran 12 weeks of protein measurements from 42 patients against six aging clock models built independently by teams at Harvard, Oxford, Peking University and Insilico. All six clocks showed reduced biological markers for age, while patients on placebo held steady or edged higher. Insilico presented the work as a framework for measuring aging inside conventional disease trials, in line with FDA biomarker guidance. [Bloomberg](https://www.bloomberg.com/news/articles/2026-09-07/ai-discovered-drug-reverses-aging-markers-in-study-biotech-says) [New York Times](https://www.nytimes.com/2026/09/07/science/ai-generated-drug-longevity.html) - **OpenAI tests ChatGPT setting that reads users' Gmail, Slack and Drive to copy their writing** — OpenAI confirmed it is testing a "Writing Style" option that has ChatGPT draw on a user's existing messages and documents rather than on samples pasted into a conversation, per BleepingComputer. An onboarding screen shown to a small group of testers says ChatGPT "will write in your voice by referencing examples from your connected apps." The company lists Slack for messaging, Google Drive and Notion for documents and Gmail for email as the services it can draw on. Anthropic's comparable Styles feature required users to upload or paste writing samples themselves. [BleepingComputer](https://www.bleepingcomputer.com/news/artificial-intelligence/chatgpt-can-now-connect-to-your-personal-apps-to-mimic-writing-style/) # Industry & Market Watch - **ASML lines up chipmakers to build largest data center chips** — ASML said it will work with major customers to move its High NA extreme ultraviolet machines, its newest chip printing tools, to a 12 inch mask, per Reuters. The mask is the stencil carrying a chip's pattern, and a larger one would lift the size limit that stops those tools from printing full size data center processors. Existing EUV machines top out at roughly 800 square millimeters. The company plans a pilot line in 2031 and production readiness in 2033, and said the shift would raise those systems' productivity by 40%. TSMC said it will use High NA in high volume production from 2030, and Samsung plans to bring it into memory chip production by 2028. [Bloomberg](https://www.bloomberg.com/news/articles/2026-09-08/asml-tsmc-samsung-intel-back-12-inch-masks-for-ai-chips) [Reuters](https://www.reuters.com/world/asia-pacific/asml-work-with-major-chipmakers-to-use-latest-tools-larger-chips-2026-09-08/) - **Mistral raises about $3.5 billion led by Samsung, nearly doubling valuation to about $24 billion** — The French model developer closed a Series D of 3 billion euros ($3.5 billion) led by Samsung Electronics, lifting its valuation to 21 billion euros (about $24 billion) from 11.7 billion euros (about $14 billion) a year ago, per Le Monde. The European Commission's Scale-up Europe fund and existing shareholder PSG Equity joined the round. It follows criticism since the summer that Mistral has drifted from building frontier models toward selling computing capacity and hosting other developers' systems, including one from the Chinese firm Z.ai. Chief Executive Arthur Mensch said the computing power the company owns will roughly double over five years, on top of the 4 billion euros (about $4.7 billion) it is already sinking into data centers in France and Sweden. [Le Monde](https://www.lemonde.fr/en/economy/article/2026/09/08/mistral-ai-raises-3-billion-in-response-to-doubts-over-its-strategic-direction_6757278_19.html) [Euronews](https://fr.euronews.com/business/2026/09/08/mistral-ai-leve-un-montant-record-de-3-milliards-deuros-lors-dun-tour-mene-par-samsung) - **Bankers seek investment grade ratings for OpenAI and Anthropic ahead of listings** — Bankers working on the planned initial public offerings of OpenAI and Anthropic are pressing credit rating agencies to award the two labs investment grade ratings, per the Financial Times. Morgan Stanley and Goldman Sachs recently held talks with the agencies on behalf of both companies. An investment grade designation would lower borrowing costs for the labs and for the infrastructure partners building data centers around them. [Financial Times](https://www.ft.com/content/aa304856-cade-4ad8-a2bf-2dd34fa75b1b?syn-25a6b1a6=1) [Boursorama](https://www.boursorama.com/bourse/actualites/selon-le-financial-times-les-banquiers-d-openai-et-d-anthropic-cherchent-a-obtenir-des-notations-de-credit-investment-grade-apres-leur-introduction-en-bourse-381f9ac3808fc9373aeb1360ed18ed9d) - **Anthropic drops $6 billion Decart takeover after due diligence** — Anthropic has decided against buying Decart AI, the Nvidia backed startup behind the Lucy live video editing model and the Oasis simulation platform, people familiar with the matter told Bloomberg. The talks, priced at about $6 billion and noted in AIPD's [August 13th edition](https://aipolicydaily.org/archive/daily/2026-08-13/), ended after Anthropic completed due diligence. The two companies may still pursue other forms of collaboration. [Bloomberg](https://www.bloomberg.com/news/articles/2026-09-08/anthropic-said-to-walk-away-from-6-billion-decart-acquisition) [The Economic Times](https://m.economictimes.com/tech/artificial-intelligence/anthropic-drops-pursuit-of-ai-startup-decart/amp_articleshow/133906210.cms) # Global & Geopolitics - **Aria chair Matt Clifford quits over Anthropic job after MPs flag conflict of interest** — Matt Clifford stood down as chair of the Advanced Research and Invention Agency (Aria), the UK unit that funds high risk science and technology projects, after taking a full time job at Anthropic leading its engagement with governments outside the U.S., per the Guardian. Labour MP Chi Onwurah, who chairs the Commons science, innovation and technology committee, had called the arrangement a "clear conflict of interest" in a letter to the AI minister, Kanishka Narayan. Anthropic had said Clifford would recuse himself from any Anthropic business at Aria under terms agreed with the government. He said he was leaving so the new role would not become a distraction from Aria's work, and will remain in the job until Nov. 6. [The Guardian](https://www.theguardian.com/technology/2026/sep/07/architect-uk-ai-policy-quits-anthropic-conflict-of-interest-concerns) - **Macron presses von der Leyen for EU wide social media ban for under 15s** — President Emmanuel Macron asked European Commission President Ursula von der Leyen to bring forward EU legislation barring children under 15 from social media platforms, in an Aug. 29 letter his office disclosed, per Reuters. France's Constitutional Council struck down a national version of the ban this summer, before it was due to take effect. Von der Leyen said in July that the bloc would limit young children's access through a tiered system giving under 13s only limited supervised access. That is a softer approach than the blanket cutoff Macron wants. Macron said France will separately rework the bill the court rejected. [Reuters](https://www.reuters.com/legal/litigation/macron-urges-eu-ban-social-media-under-15s-letter-von-der-leyen-2026-09-08/) [Bloomberg](https://www.bloomberg.com/news/articles/2026-09-08/macron-demands-eu-action-to-ban-social-media-for-kids-under-15) - **Australia proposes letting users switch off algorithmic feeds, with AI chatbots inside the duty of care** — Australia's government released draft Digital Duty of Care legislation that would require platforms to notify every account holder of a feed choice, per AP. Those over 16 could then choose between an algorithmically recommended feed and one limited to accounts they follow. Penalties for platforms that fail to reduce foreseeable harm would run to 109.2 million Australian dollars ($78.6 million). The duty reaches beyond social media to online games, apps and AI chatbots, which would have to shield children from design features that encourage addiction or damage self-esteem. Prime Minister Anthony Albanese said the scheme, branded "My Feed, My Way," puts choice back in users' hands. The draft is out for consultation, and the Greens and the opposition Coalition said they will work through the detail before settling a position. [AP](https://apnews.com/article/1eee351b39f22b340ef289c792f7993b) [Bloomberg](https://www.bloomberg.com/news/articles/2026-09-08/australia-sets-up-new-big-tech-fight-with-bid-to-block-targeting) [ABC News](https://www.abc.net.au/news/2026-09-08/labour-s-social-media-algorithm-choice-duty-of-care-bill/107130100)