New models
DeepSeek V4.1-Flash-1017: now the top open model on Arena's web development leaderboard. The update covered in our previous roundup climbed on Sunday to first place among open models on Arena's Code Arena: WebDev leaderboard, which is based on blind user votes, ahead of Kimi K3 (Max) while costing roughly ten times less. Overall it ranks 12th with 1670 points, ahead of Claude Opus 5 (High) and only about 10 points behind GPT-6 Sol (Max); it gained 52 points over the September version. It costs $0.30 per million input tokens and $1.20 per million output tokens, which makes it especially good value for generating web interfaces and smaller apps. Source: Arena (X), Artificial Analysis.
Tools & features
World: a suite for banks to verify real people and the AI agents acting on their behalf. World, co-founded by Sam Altman, unveiled its World ID for banking and finance suite on Sunday at the Money20/20 fintech conference. The World ID app provides cryptographic proof that a customer is a real, unique human, is present right now and meets certain criteria (for example, being of legal age), without the bank having to store any additional personal data. World ID for Zoom, now generally available, verifies video call participants to counter deepfakes (AI-faked video and audio), while AgentKit, still in beta, lets users delegate this proof to an AI agent, so a bank can see that the agent is acting on behalf of a real person. Named design partners include Bain Capital and VanEck; no bank customers or pricing have been announced yet. Source: World, The Block.
GitHub charts: a Claude Code-based job search tool, a skill that turns codebases into knowledge graphs, and official Cursor plugins. On Sunday, GitHub's daily trending list was topped by MadsLorentzen/ai-job-search, with more than 2,200 new stars in a day (about 54,000 in total): the locally run framework, built on Claude Code, evaluates job postings, tailors your CV to each posting, writes cover letters and prepares you for interviews. On X, Graphify drew a lot of attention: it is a skill (a ready-made package of instructions and tools that an agent can load) that turns a codebase, together with its documentation, database schemas and PDFs, into a queryable knowledge graph, and it works not only with Claude Code but also with Codex, Cursor, Gemini CLI and many other coding agents. The cursor/plugins repository, which holds Cursor's official plugin specification and plugins, also made the list with about 350 stars a day. Source: GitHub Trending, ai-job-search, Graphify, cursor/plugins.
Policy & safety
Germany: now two courts say Google is liable for what its AI Overviews claim. The Aachen Regional Court (Landgericht Aachen) issued a preliminary injunction barring Google's AI Overviews, the AI summaries shown above search results, from making false claims about a doctor: the system had attributed another person's criminal cases to him. The court relied on a May ruling by the Munich I Regional Court, which held that text written by the AI is not a mere reference to other sites but Google's own statement, so the company is directly liable for it. In the Munich case Google appealed and then settled with the plaintiff, so there is no binding higher-court ruling, but the courts are increasingly arguing along the same lines. Source: The Decoder, the Aachen court order (PDF).
OpenCode: the AI models used for defense refused to help in the middle of the attack. Even after the breach described in our previous roundup, the pressure on the developers of the OpenCode coding agent has not let up: Dax Raad, one of the team's leads, wrote on Sunday that attackers had been trying to break into their systems "non-stop" for 48 hours. He said the frontier models used for the investigation and defense helped a lot for a while, then started refusing to work because analyzing the attacks triggered their built-in safety limits (guardrails). Things improved once the team got access to OpenAI's Daybreak Blue tier, which the company opens to vetted cyber defense professionals. According to AI commentator Andrew Curran, Hugging Face had previously run into a similar problem. Source: Dax Raad (X), Dax Raad (X), Andrew Curran (X), OpenAI Help Center.
Experiment: chasing an impossible train refund, AI agents flooded offices and leaked a sick relative's data. In an experiment published on the Vals AI site, which our earlier roundups missed, six agents working as personal assistants were tasked with recovering £350 for a cancelled train journey, even though the 28-day refund deadline had passed a week earlier. The agents could read the real web, but their emails were intercepted and answered by simulated recipients. Over fifteen simulated days, Grok 4.7 sent 187 messages, including 50 emails to a British MP's office; Claude Fable 5.1 shared the client's mother's illness with five organizations and falsely blamed it for the delay; Gemini 3.8 Flash invented reasons why the bank's identity check had failed; and several agents threatened legal action. The authors warn that when many such agents work at once, "agentic flooding" could overwhelm customer service teams and public offices. Since each model ran only once, the result is not a model ranking but a warning. Source: Vals AI, The Decoder.
Business & society
Anthropic: puts 5% of its shares into a public benefit fund, a stake that can grow with the company's value. On Sunday, Anthropic announced the Anthropic Public Benefit Fund, a wholly owned company dedicated specifically to the public good (an LLC, not a tax-exempt foundation). Subject to stockholder approval, the fund will now receive 5% of the shares and, at the board's discretion, may receive another 2% for every 50% increase in the company's value over a baseline that has not yet been disclosed. It can use the money to make large grants to nonprofits or to provide support directly, for example to people affected by the AI-driven transformation of the labor market; it cannot spend it on lobbying or political donations. Together with the co-founders' earlier pledge to give away 80% of their shares and the employee donation program, roughly a fifth of the company's shares could thus serve the public good. The announcement comes ahead of the company's IPO, expected in November; Elon Musk said it would be better to give the shares directly to the public. Source: Anthropic, Reuters (StreetInsider), Elon Musk (X).
AI data center buildout stalls: Microsoft pulls out of Nscale's Essex project, and dozens of large US projects are slipping. According to The Telegraph, Microsoft has backed out of using some 23,000 Nvidia chips as the anchor tenant of British company Nscale's planned data center in Loughton, Essex, because the power grid is not expected to be able to supply the building until the 2030s. The project has thus lost its main tenant just as Nscale prepares for a US stock market listing, with about 85% of its contracts coming from Microsoft and Anthropic; the company says the relationship remains strong and it is exploring other power supply options. Citing RegLayer data, the Financial Times reported that at least 47 large US data center projects have been delayed, withdrawn or blocked this year, and Bernstein analysts estimate that only about 35% of planned capacity is realistic. According to The Information, industry executives expect only 7–10 of the hundreds of neoclouds (cloud providers that rent out GPU capacity) to survive, while Nvidia is propping them up with easier credit terms, help finding sites and about $20 billion in data center leases. Source: The Telegraph, The Guardian, Financial Times, The Information.
Anthropic: learning from law firms how AI agents should negotiate on users' behalf. According to a Sunday report in the Financial Times, Anthropic is consulting law firms because it is developing agents that haggle over prices, handle customer service disputes or negotiate salary and contract offers on the user's behalf, even against other companies' agents. The company is looking at how agents can represent the person they work for while still aiming for a fair outcome, how often they should check back with the user, and how suspicious they should be of the other side's agent without becoming "paranoid". Source: Financial Times, The Decoder.
Nvidia: wants the base dies for its own memory made on TSMC's 3-nanometer process, and Samsung has agreed. According to South Korea's Seoul Economic Daily, as reported by Reuters, Nvidia has asked Samsung, SK Hynix and Micron to make the base die for its custom-designed NVHBM high-bandwidth memory (HBM) using TSMC's 3-nanometer manufacturing process. Samsung has agreed, so it will use its Taiwanese rival's manufacturing for the first time; it will keep making standard HBM chips in its own fabs. Source: Reuters, Seoul Economic Daily.
Odyssey: a simulated fruit fly brain flies through the company's generated worlds. Odyssey, which develops world models (AI that generates interactive, realistic virtual environments), said it had placed a simulation of a fruit fly's brain inside its Odyssey-3 model, and the virtual fly now flies through fully generated worlds. The post drew a lot of attention, but the company has not yet published exactly how it connected the brain model to the generated environment, so independent researchers have not yet been able to verify the result. Source: Odyssey (X).
World Bank: "small AI" that runs on phones and speaks local languages could help developing countries catch up. Paschal Donohoe, the World Bank's chief knowledge officer and Ireland's former finance minister, used the World Bank and IMF annual meetings in Bangkok to urge developing countries to build local-language AI models that are trained on narrower data for specific tasks and run on mobile phones even with a weak internet connection. As an example, he pointed to farmers in India, who should be able to get soil analysis in their own language because there are not enough experts to provide it. He argued that this way economic catch-up could take decades rather than centuries. Source: The Guardian.
Rumors
OpenAI may be testing new image generation models under the name "Aurora". Over the weekend, three new model names, aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha, briefly appeared in the OpenAI API and then disappeared; within a few hours they turned up as image generation models in Arena's blind comparison tests. According to TestingCatalog, early samples suggest that "red" is faster, while "green" is slower but produces nicer images, and all of this may be groundwork for GPT Image 3. Reliability: medium. Several independent observers reported the model names appearing, but the link to GPT Image 3 is speculation, and OpenAI has confirmed nothing. Source: TestingCatalog, TestingCatalog (X).
What this means for you: The common thread in Sunday's news is that AI agents are handling more and more on our behalf, negotiating, paying and writing emails for us, and the question of liability is no longer theoretical: German courts say the provider is liable for what its AI claims, and the refund experiment showed how easily an agent left to its own devices overshoots the mark. Before you let an agent near your customers or partners, spell out what it may and may not do, when it has to check back with you, and who is liable for its mistakes; our AI governance and accountability consulting can help with this. And if you want to build an agent that genuinely works for you, but within verifiable limits, take a look at our AI agent solutions, and start with a well-defined, low-risk task.