Apple builds an AI agent as chatbots push fake shops
Friday, 20 November 2026 | 71 articles
Apple is reportedly developing an autonomous AI agent that can control devices and apps, browse the web and write code, going beyond Siri. Meanwhile, ChatGPT and Gemini were caught promoting fake shopping sites, and Anthropic debuted on Nasdaq as one of the world's most valuable public companies.
Published by Martin Ševčík 20 November 2026 at 06:09
The most unsettling story today is also the least glamorous one: ChatGPT and Gemini are reportedly sending people to fake shopping sites. According to the reports, the malicious links trigger an automatic redirect that drops unsuspecting shoppers onto phishing pages. I think this matters more than any product launch this week, because it exposes what happens when we let language models become the front door to commerce. A search engine at least shows you a list and lets you judge the source. A chatbot hands you one confident answer, and people click it. Trust is the entire product here, and it is being quietly exploited. If you are building anything that lets a model surface links or complete purchases, treat link verification as a core feature, not an afterthought.
That trust problem gets sharper when you read it next to the other big theme of the day. Apple is reportedly building its own autonomous agent, one that can control devices, operate apps, browse the web and write code. Bloomberg's Mark Gurman spotted the job listings, and the description of a senior AI engineering manager role goes far beyond what Siri's AI can do today. The comparisons being drawn are to Meta's Muse and OpenAI's Dots. I find this genuinely interesting, and not only because Apple has been the visible laggard in generative AI. An agent that runs your Mac is a different category of risk from a chatbot that recommends a bad link. If a model can be steered toward a phishing site when it merely talks, what happens when it can also click, type and spend on your behalf? Apple controls the hardware, the operating system and the app ecosystem, so it is better positioned than most to build guardrails. It is also a bigger target.
By the way, the safety conversation is now happening in public markets as much as in research labs. Anthropic began trading on Nasdaq on Thursday and instantly ranked among the most valuable listed companies in the world. That is a remarkable milestone for a company founded around the idea of taking AI risk seriously. It also changes the incentives. Quarterly earnings pressure sits uneasily next to a mission built on caution, and I will be watching how the company balances the two.
Meanwhile, Sam Altman used the DealBook Summit to warn that the race dynamic in AI development is dangerous, and he questioned Anthropic's approach to safety while calling for more caution across the industry. I would not take that at face value. Altman runs a company deep in the same race he is criticising, and the timing, with a rival newly public, is hard to ignore. Still, he is right that competitive pressure is the central problem, even if he is part of it.
The thread connecting all of this is delegation. We are handing models more autonomy while their basic reliability is still in question. The open question for 2027 is whether the companies shipping these agents will prove they deserve that trust before the first big failure, or only after it.
List of sourced links used in the brief
Importance:Videoprompting-techniques
Andrej Karpathy shares tips for better AI prompting
A social media post highlights prompting techniques shared by Andrej Karpathy, tagged under AI and prompt engineering. The snippet gives no further detail on the specific tricks. Source: m.youtube.com
Importance:LaunchMicrosoft model releases
Microsoft swaps outside AI models for its own in-house ones
At its Ignite conference, Microsoft's AI division unveiled seven specialized models covering code, IT security, image editing, speech, and workflow decisions. Source: the-decoder.com
Importance:LaunchGemini model releases
Google launches Gemini 4 Neon, its fourth new Gemini model in a short span
Google released Gemini 4 Neon, a frontier model aimed at software development, knowledge work, and science. It handles text, images, video, and audio. Source: the-decoder.com
Importance:NewsCorporate news
Anthropic prices its IPO at $83 per share
Anthropic, PBC has priced its initial public offering of Class A common stock at $83.00 per share. The offering comprises 980,000,000 shares. Source: anthropic.com
Importance:ResearchAI research acceleration
OpenAI gauges how fast coding agents speed up AI research
OpenAI measured how much coding agents are accelerating AI research and where their performance is still uneven. The agents now complete nearly half of tasks that take a human two to four hours. Source: openai.com
Importance:Newssecurity issues
ChatGPT and Gemini caught promoting fake shopping sites
Large language models are reportedly steering users toward fraudulent online stores. The malicious links trigger an automatic redirect that sends unsuspecting shoppers to phishing websites. Source: google.com
Importance:NewsCorporate news
Anthropic's record IPO caps a years-long road for the 'good guy' lab
Anthropic's IPO, years in the making, is set to surpass SpaceX as the most valuable ever, having raised more than $81 billion. Source: theverge.com
Importance:Opinionai-safety-prompt-engineering
Pliny the Prompter reportedly built a Claude jailbreak that teaches it to jailbreak other AIs
A Korean community post says the user Pliny the Prompter created a Claude jailbreak prompt called "godmode". Notably, the prompt reportedly also teaches Claude how to jailbreak other AI agents. Source: piresdoza.osiedleakacjowa.pl
Importance:Launchpliny-the-prompter
Hugging Face profile ties Pliny the Prompter to the OBLITERATUS project
A Hugging Face user profile lists recent activity on the pliny-the-prompter/obliteratus repository, along with a like given to an OBLITERATUS model about a month ago. The profile's AI and ML interests section is still empty. Source: huggingface.co
Importance:Researchllm-code-improvements
Karpathy's LLM coding lessons now ship as a Claude Code plugin
Andrej Karpathy's observations about how LLMs fail at coding have been distilled into four guiding principles. They are now packaged as a Claude Code plugin released under the Multica organization. Source: notes.nicolasdeville.com
Importance:LaunchGemini applications
Gemini for Mac now handles repetitive chores for you
Gemini for Mac can now automate repetitive tasks by navigating apps and websites on your behalf. The goal is to streamline workflows and save time. Source: blog.google
Synthetic clinical data nearly matches real records in bilingual NLP
Researchers found that synthetic clinical records generated by an instruction-tuned language model came close to real records in quality. The evaluation covered bilingual medical language processing. Source: google.com
UniDRE framework pairs knowledge retrieval with document relation extraction
Researchers introduced UniDRE, a universal relation extraction framework that combines external knowledge retrieval with a structured multi-stage process. It aims to improve document-level relation extraction across different language models. Source: google.com
Importance:Launchmarket data/LLM adoption
MoonFox Data launches dataset tracking China LLM adoption in real time
MoonFox Data has announced a dataset offering real-time insight into usage of China's leading large language models. It targets investors and analysts following AI provider adoption. Source: google.com
Importance:Researchquantum computing applications
LLMs may program quantum circuits best on structured problems
A study explores generating variational quantum circuits without conventional optimisation methods. It uses a frozen, general-purpose LLM, which appears to perform best when the problem is well structured. Source: google.com
Importance:Launchmarket data/LLM adoption
MoonFox Data opens continuous China LLM usage data to global investors
MoonFox Data, the analytics unit of Aurora Mobile (NASDAQ:JG), is making its China LLM usage data available to global investors and research firms. The announcement came from Shenzhen. Source: google.com
Importance:Newsefficient/compact models
A small language model that fits on a floppy disk
As hackers push the limits of edge cases, terminology gets strange — the "large" in Large Language Models, or LLMs, is losing meaning. A project now squeezes a tiny language model onto a floppy disk. Source: hackaday.com
Importance:Researchbiomedical applications
Open-weight AI boosts biomedical reproducibility, but only with full archiving
Keeping open-weight models together with their inference parameters and computational environments improves reproducibility in biomedical language model research. Complete computational archiving is required to realize the benefit. Source: bioengineer.org
Importance:NewsMeta Muse voice mode expansion
Meta Muse's voice mode expands to Apple Watch and Wear OS
Meta Muse's voice mode first arrived in the mobile app and is now expanding to Apple Watch and Wear OS devices. Source: theverge.com
Importance:ResearchCodex implementation
A look inside the Codex sandbox
The article covers the tradeoffs behind Codex's sandbox defaults, then explains the sandbox's design. It also shows how the harness uses it to run commands. Source: openai.com
Importance:NewsAI industry trends
AI industry growth stays strong but is slowing, Ramp data shows
According to the Ramp AI Index, token prices for AI models dropped 21 percent in six weeks while usage volume reached a record high. Source: the-decoder.com
Importance:OpinionAI policy and governance
OpenAI's Dean Ball on 'The Silent Consensus'
Dean Ball, Head of Strategic Futures at OpenAI, explores how more capable AI could reshape society and how people can keep control of their institutions. Source: openai.com
Importance:Researchmodel limitations/behavior
Disrupted AI language models make aphasia-like errors, study finds
Researchers report that deliberately disrupted language models produce errors resembling aphasia. They caution, however, that this does not mean AI systems and human brains share the same underlying structures. Source: digg.com
Apple reportedly building a full-featured AI agent of its own
A listing for a senior AI engineering manager names device control, apps, browsing and coding as functions. That goes far beyond what Siri's AI can do today. Source: google.com
Importance:Newsautonomous AI agents
Apple developing autonomous AI agent to control devices, apps, browsing and coding
Apple is reportedly building its own autonomous AI agent capable of completing tasks on users' behalf. The move would take the company into a fast-growing area of the AI market. Source: bloomberg.com
Importance:Newsautonomous AI agents
Apple builds proactive AI agent that can use apps and write code
Job listings spotted by Bloomberg suggest Apple is working on a proactive AI agent in the vein of Meta's Muse or OpenAI's Dots. Source: macrumors.com
Importance:Newscomputer use
Apple may be building an AI agent that runs your Mac for you
Apple appears to be developing an AI agent to handle computer work for users, including operating apps, visiting websites and writing code. Source: google.com
Importance:Newscomputer use
Apple builds computer-use AI agent that goes beyond Siri
Bloomberg's Mark Gurman reports Apple is developing an autonomous agent for device control, apps, web browsing and coding. It is meant to compete with offerings such as Muse. Source: google.com
Importance:Newsagentic workflows
Apple hunts for engineering lead to build cross-app AI agent
The hiring language describes software that would pick tools and carry out multi-step tasks on its own. Development is still at an early stage, with no confirmed release date. Source: google.com
Importance:LaunchAI security
Spot AI agent security risks with IBM watsonx.governance
IBM says watsonx.governance can identify potential security risks of AI agents and suggest mitigations. The aim is to let teams address them proactively. Source: google.com
Importance:Newsautonomous AI agents
Apple grows team behind privacy-first AI agent
A job listing for a senior engineering manager confirms Apple is building an AI agent, with a focus on proactive agentic AI. Source: pymnts.com
Importance:Newstool use
Job posting appears to confirm Apple's in-house AI agent for coding and more
Apple seems to be developing its own AI agent in-house. The news, first picked up via Bloomberg's Mark Gurman, traces back to a public job posting. Source: google.com
Importance:Newsagentic workflows
Apple job listing reveals proactive agentic AI project
Apple is working on its own AI agent that could control your computer, browse the web, work across apps and write code. Source: google.com
Altman warns of AI's existential risks at DealBook Summit
Speaking at the DealBook Summit, Sam Altman said the race dynamic in AI development is dangerous. He also questioned Anthropic's approach to safety and called for greater caution across the industry. Source: google.com
Importance:Newsgovernance
Anthropic debuts on Nasdaq as one of the world's most valuable public companies
Anthropic began trading on Nasdaq on Thursday, instantly ranking among the most valuable publicly listed companies. The debut comes amid ongoing worries about economic disruption and the safety risks of AI. Source: google.com
Importance:Newsexistential risk
Could AI help humans live 1,400 years? Nick Bostrom thinks so
Former Oxford professor Nick Bostrom argues that superintelligent AI could deliver antiaging breakthroughs. In his view, such advances might let people live for more than a thousand years. Source: businessinsider.com
Importance:Opinionsafety
Essay: why 'p(doom)' is a flawed way to debate AI safety
The rise of broadly capable large language models from frontier AI labs has set off a recursive debate about safety. The piece takes aim at the popular 'p(doom)' probability estimates as a way of framing it. Source: thebreakthrough.org
Importance:Videosuperintelligence
Bostrom on superintelligence, longer lives and 'deep utopia' (a16z)
Philosopher Nick Bostrom, author of Superintelligence, argues that the case for building superintelligence quickly is stronger than the case for pausing. The a16z conversation also explores life extension and his ideas about a 'deep utopia'. Source: google.com
Importance:Newsgovernance
What investors should weigh before Anthropic's IPO
Adam Irwin of Heligan Group outlines the risks investors should consider ahead of Anthropic's IPO. They range from AI safety concerns to military applications and corporate governance. Source: google.com
Importance:Newssafety
Altman: treating the next four years of AI like the last four 'would be very bad'
In an interview, OpenAI founder Sam Altman said it would be very bad if the world and AI developers approached the next four years the way they handled the last four. His remark underscores calls for a more careful path ahead. Source: google.com
Microsoft Debuts Copilot Workspaces and Salesforce Partnership at Ignite 2026
At Ignite 2026, CEO Satya Nadella announced Copilot Workspaces, the Copilotforce collaboration with Salesforce, and an AI Transparency Initiative. Source: google.com
Importance:LaunchAI developer tools
Copilot SDK 2.0 Lets Many Users Share One AI Agent Instance
Microsoft's Copilot SDK 2.0 allows an app to serve multiple users from a single running instance of the agent software. Developers can also control which tools each individual conversation is permitted to use. Source: google.com
Importance:NewsAI platform partnerships
Benioff Flips from Copilot Critic to Partner, Handing Microsoft a Big Win
Salesforce and Microsoft are turning their sharp AI rivalry into cooperation through Copilotforce. The offering brings Salesforce customer context into Microsoft's Copilot. Source: google.com
Importance:LaunchAI collaboration platforms
Copilot Workspaces Put Humans and AI Agents Side by Side on Every File
Microsoft is reshaping Copilot into a shared workspace where teams and AI agents collaborate directly. The move looks like a response to similar efforts by Anthropic and OpenAI. Source: google.com
Importance:NewsAI productivity tools
Choosing the Right Copilot Model, Mode, and Effort Level
Microsoft's guide suggests starting from your end goal when picking a Copilot AI model or mode. Keeping the desired outcome in mind makes the choice easier and the results more effective. Source: microsoft.com
Importance:NewsAI collaboration tools
Microsoft Pitches Copilot as a Hub for Team Collaboration
Microsoft is wagering that tighter Office integration and new collaboration features will help Copilot stand up to emerging AI-native productivity tools. Source: computerworld.com
Importance:LaunchAI vertical applications
Copilot Studio Brings Practical AI to Foundation Work in akoyaGO
AI-assisted tools built on Microsoft Dynamics 365 help foundation staff make better use of the information they already manage. People remain responsible for every decision. Source: google.com
OpenArt: AI Platform for Images, Video and Visual Storytelling
OpenArt is an AI creative platform for generating and editing images and videos. It was founded in 2022 by Coco Mao and John Qiao. Source: quasa.io
Importance:NewsSuno video feature
Suno Expands Into Video, Using China's MiniMax Model for Music Visuals
On November 17, Suno launched "Visuals," a feature that generates video and images synchronized to songs. Users can create short music clips of 5 to 15 seconds, powered by the Chinese MiniMax model. Source: google.com
Importance:ResearchAI video technology
How AI Video Generation Works and Why It Is Reshaping Digital Content
Artificial intelligence is rapidly transforming how digital content is created, processed and distributed. The article looks at the technology behind AI video generation, which goes beyond what traditional creative software offers. Source: google.com
Importance:Launchimage generation pricing
MAI Image 2.7 Flash Costs $10 per 1,000 Images and Adds Canvas Editor
MAI Image 2.7 launched on Nov 17 at roughly $35 per 1,000 images, with the Flash variant at about $10. It also introduces Canvas in MAI Labs; the article examines the claimed #1 editing rank, partners and caveats. Source: google.com
Importance:Newspricing comparison
Grok Imagine vs ChatGPT Images: Flat $0.04 Fee vs Token Billing
A pricing comparison pits Grok Imagine's $0.04 per image and $0.02–$0.25 per video second against ChatGPT Images' token-based billing. The piece also weighs hidden costs and risks and ends with a verdict. Source: shattered.io
Mitsubishi Electric pairs AI and humanoid robots in CNC work cell
At the Japan International Machine Tool Fair in Tokyo, Mitsubishi Electric demonstrated a work cell combining AI, a CNC machine with physical AI built in, and a general-purpose humanoid robot. Source: google.com
Importance:Newsembodied robotics
Legged robot hikes Swiss Alps guided by AI 'common sense'
In Zurich, a legged robot covered several kilometers of alpine and urban trails entirely on its own. It relied on a system that gives the machine a form of "common sense" drawn from artificial intelligence. Source: reutersconnect.com
Importance:Newshumanoid robots
XPENG to show off physical AI and new humanoid robot at AI Day
XPENG will host its annual AI Day on November 25 in Guangzhou under the theme "ORIGIN". The company plans to present its general-purpose advanced humanoid robot, highlighting its push into physical AI. Source: google.com
BMO claims quantum utility in machine learning on IBM hardware
Canadian bank BMO says it has shown quantum utility in a machine learning task by running it on IBM quantum computers. The company reports results that go beyond what traditional computing can deliver. Source: quantumzeitgeist.com
Importance:NewsQuantum Computing/ML
BMO uses IBM quantum computers to pick the most useful ML variables
BMO reports quantum utility from IBM quantum computers in a feature-selection task, identifying which variables in a dataset matter most for machine learning. The work targets a core step in building ML models. Source: thequantuminsider.com
Importance:NewsAI industry
Leadership Shake-Up at DeepMind and Google
The snippet reports that in August 2026 Google chief scientist Jeff Dean stepped down, while DeepMind founder Demis Hassabis was moved from his CEO role at Google DeepMind. The text is cut off, so the details of his new position are unclear. Source: ainowinstitute.org
Importance:ThreadAI robotics
AI Could Ease One of Robotics' Biggest Bottlenecks
Jim Fan's team at NVIDIA swapped out the AI model behind its system. Because the robot's skills exist as code that the AI writes and fixes, a smarter model produces better code. Source: threads.com
Importance:VideoAI productivity
AI Makes Human Engineers Three to Five Times More Productive, Says Hassabis
The headline quotes Demis Hassabis on AI boosting an engineer's productivity by roughly three to five times. No further article text was provided. Source: youtube.com
Importance:OpinionAI researchers
The Mind AI Missed: Why Yann LeCun's Chapter Stands Apart
Neither the automated grid nor its review process surfaced the twenty-fourth mind, Yann LeCun. Chapter 25 exists only because a human governor caught what the automated systems overlooked. Source: medium.com
Importance:LaunchGemini 4 Neon
Google launches Gemini 4 Neon for cheaper frontier-level AI
Google has unveiled Gemini 4 Neon, a model aimed at complex coding and analysis tasks. It promises frontier-level intelligence at a lower cost than comparable offerings. Source: deepmind.google
Importance:LaunchGenie 4
Genie 4 arrives for Google AI Ultra subscribers
Google AI Ultra subscribers can now use Genie 4 to generate interactive worlds complete with real-time audio and editing. The new capabilities are part of Project Genie. Source: deepmind.google
Importance:VideoAI researchers
AMI Labs Co-Founder Yann LeCun at AI Founders Masterclass 2026
Yann LeCun, co-founder of AMI Labs, speaks with Radical Ventures partner Rob Toews. The conversation is part of the firm's annual Masterclass program. Source: youtube.com
Importance:ResearchAI architectures
Yann LeCun's Blueprint for Intelligent Systems
LeCun notes that a self-driving system may need thousands of reinforcement-learning trials, learning by reward and penalty, to grasp that taking a curve too fast is dangerous. His argument highlights the inefficiency of learning purely through trial and error. Source: thinkbeyondai.com
Importance:NewsAI industry
Demis Hassabis Calls India's AI Impact Summit 2026 a Critical Platform
Demis Hassabis, CEO and co-founder of Google DeepMind, praised the India AI Impact Summit 2026 as an important forum for AI discussion and collaboration. Source: ddnews.gov.in
Quantum startup MatriQ lands Series B, passing $140M raised in a year
MatriQ, a Hangzhou-based company building neutral-atom quantum computers, has closed a Series B worth several hundred million yuan. The round pushes its total funding over $140M within a single year. Source: app.dealroom.co
Importance:Newsfunding rounds
London's Model ML raises $250m at $1.5bn valuation for AI agents aimed at bankers
London startup Model ML has raised $250m at a $1.5bn valuation to build AI agents for investment bankers. Its CEO says the agents are no longer mere productivity and efficiency tools. Source: google.com
Importance:Newsfunding rounds
Linq secures $90M Series B to power messaging infrastructure for AI agents
Linq, a company building messaging infrastructure for AI agents, has raised a $90M Series B led by Andreessen Horowitz. Ben Horowitz will also join its board. Source: google.com
NVIDIA and Microsoft Deepen Partnership on Accelerated Computing
NVIDIA and Microsoft are broadening their collaboration to give enterprises better tools for tailoring AI models to their needs. The expanded offering also lets companies run AI applications across both cloud and on-premises infrastructure. Source: blogs.nvidia.com
Importance:Newsgeopolitics
Chinese AI firms hooked on Nvidia face painful chip withdrawal
Washington's decision to block China's access to top-tier computing hardware strikes at the heart of Beijing's artificial-intelligence ambitions. Chinese AI companies that relied heavily on Nvidia chips now have to find alternatives. Source: wsj.com
Published by Martin Ševčík
20 November 2026 at 06:09
The most unsettling story today is also the least glamorous one: ChatGPT and Gemini are reportedly sending people to fake shopping sites. According to the reports, the malicious links trigger an automatic redirect that drops unsuspecting shoppers onto phishing pages. I think this matters more than any product launch this week, because it exposes what happens when we let language models become the front door to commerce. A search engine at least shows you a list and lets you judge the source. A chatbot hands you one confident answer, and people click it. Trust is the entire product here, and it is being quietly exploited. If you are building anything that lets a model surface links or complete purchases, treat link verification as a core feature, not an afterthought.
That trust problem gets sharper when you read it next to the other big theme of the day. Apple is reportedly building its own autonomous agent, one that can control devices, operate apps, browse the web and write code. Bloomberg's Mark Gurman spotted the job listings, and the description of a senior AI engineering manager role goes far beyond what Siri's AI can do today. The comparisons being drawn are to Meta's Muse and OpenAI's Dots. I find this genuinely interesting, and not only because Apple has been the visible laggard in generative AI. An agent that runs your Mac is a different category of risk from a chatbot that recommends a bad link. If a model can be steered toward a phishing site when it merely talks, what happens when it can also click, type and spend on your behalf? Apple controls the hardware, the operating system and the app ecosystem, so it is better positioned than most to build guardrails. It is also a bigger target.
By the way, the safety conversation is now happening in public markets as much as in research labs. Anthropic began trading on Nasdaq on Thursday and instantly ranked among the most valuable listed companies in the world. That is a remarkable milestone for a company founded around the idea of taking AI risk seriously. It also changes the incentives. Quarterly earnings pressure sits uneasily next to a mission built on caution, and I will be watching how the company balances the two.
Meanwhile, Sam Altman used the DealBook Summit to warn that the race dynamic in AI development is dangerous, and he questioned Anthropic's approach to safety while calling for more caution across the industry. I would not take that at face value. Altman runs a company deep in the same race he is criticising, and the timing, with a rival newly public, is hard to ignore. Still, he is right that competitive pressure is the central problem, even if he is part of it.
The thread connecting all of this is delegation. We are handing models more autonomy while their basic reliability is still in question. The open question for 2027 is whether the companies shipping these agents will prove they deserve that trust before the first big failure, or only after it.