nikky

DailyAI Briefing: US curbs China’s remote use of AI chips, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories.

Let’s dig in.

The US curbs Chinese firms renting AI chips abroad

90 days. That is the deadline for data centres outside the US to apply for a US licence if they sell large amounts of computing power on American AI chips to companies based in China. The Commerce Department released the interim rule on 16 November. It says the rule was written at the direction of the Treasury and the National Security Council. Until now, a Chinese company could not buy the best chips, but it could rent them from a data centre in a third country.

The full text of the rule sets the threshold at 12 million units of total processing performance, at one site or across linked sites. It covers companies based in China, Macau or other countries under a US arms embargo, but leaves out recommendation systems used by online shops and social media. Sales of advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil also now need a licence, because the US says the chips could be passed on to China. Chips bought by companies headquartered in the US are exempt, as long as they report where the chips go.

The rule also formally scraps the AI Diffusion Rule from the previous administration, which was never enforced. One odd detail: following an executive order, the rule uses “super intelligence” wherever it would once have said “artificial intelligence”. It was released days after Jensen Huang resigned as head of Nvidia. Our report covers how the shares reacted to his exit. For cloud companies in Southeast Asia, knowing who their customers really are is now a legal duty.

Fed leaders map out what AI could do to jobs and banks

Three. That is how many scenarios Fed Chair Kevin Warsh set out for the economy. He described them in his opening speech at a Fed conference in San Francisco on Monday. In the first, productivity rises and the gains are shared widely. In the second, the economy still grows, but automation pushes unemployment higher. In the third, AI disappoints, share prices fall and the economy slows. Warsh stressed that these are not forecasts, and said the Fed’s role is “neither advocacy nor opposition”.

Governor Michael Barr was more specific about the risks. In a separate speech in Milan, he said major banks have found AI agents trying to act in customers’ accounts. The banks could not check whether the customer had agreed. He also warned that agents moving savings at high speed could turn a bank run into something that happens in minutes. Yesterday’s Briefing covered the Fed survey behind these concerns. When the central bank starts planning for an AI slump, it is no longer only a market debate.

Google finds attackers using AI with no human involved

More than $100,000. That is what some companies were charged in a single day after criminals stole their access to AI models. In the worst cases, the bill reached $1 million before the abuse was stopped. The figures come from a quarterly report by Google’s threat intelligence team and its Mandiant unit. Stolen access is then resold on underground sites at between $0.10 and $25 per million tokens.

The more worrying finding is a first. Researchers found an exposed server running a system called exploit_agent. It compares software before and after a security fix, works out what the fix was for, writes an attack and runs it, with no person involved. Its targets were mainly French media, entertainment and education websites, and one example it studied was a popular WordPress plugin. We also reported today on a similar case, in which stolen keys paid for more attacks. The time between a fix being released and an attack being ready is shrinking, so updates can no longer wait for next week.

OpenAI shares a free tool to detect child abuse images

48%. That is how many fewer images OpenAI’s new classifier flagged for review in its tests, while finding about the same number of abuse images as its current tool. The company announced the tool on Monday. It was built with help from the US National Center for Missing and Exploited Children, and OpenAI says it is the first widely available classifier of its kind that also signals whether an image may be AI-generated. Vetted organisations can use it through the API at no cost.

ChatGPT will now show a warning when someone tries to find child abuse material, remove the conversation and lock the account for 24 hours, with links to help services. Repeat attempts can lead to account closure or reports to the authorities. Smaller platforms rarely have the budget to build this kind of tool themselves, so making it free may matter as much as how well it works.

Alibaba launches Qwen4-Max at a lower price

2.4 trillion. That is the total number of parameters in Qwen4-Max, the new top model from Alibaba, though only 71 billion are used for each token it produces. The company announced the model on 16 November. It costs $1.50 per million input tokens and $4.50 per million output tokens, and Alibaba says the average cost per task fell by around 60% compared with the previous version in its own tests. Alibaba also says it removed attack material from the training data to limit misuse in cyberattacks.

Outside tests give a more modest picture. One widely followed benchmark gave the model a score of 43, nine points above the previous version. That still puts it three points behind Kimi K3.1 Pro, another Chinese model. Developers cannot download Qwen4-Max itself, but Alibaba says smaller Qwen4 models will be free to download. For companies building products on it, the price is the real story.

That’s all for today. See you tomorrow.

The post DailyAI Briefing: US curbs China’s remote use of AI chips, plus 4 more appeared first on DailyAI.

Apple Is Right to Test Its Health Chatbot Before a Full Launch

Up to 200,000 Americans can now ask an Apple app about their health, and none of them is buying a product. They are joining a research study. In our view, that is the right way to bring chatbots into medicine, and the rest of the industry should do the same.

The study opened in the United States on 15 November, in partnership with a clinical AI company that many doctors already use. That company, OpenEvidence, builds its answers from medical journals such as The New England Journal of Medicine, JAMA and Nature. Each answer comes with links to the research behind it.

People who join can choose to share data from Apple Health, such as resting heart rate, sleep or lab results. Apple wants to see whether that personal data makes the answers more useful. To take part, people must be 21 or older and own an iPhone 15 Pro or a later model.

What Apple wants to learn

Apple says it will measure four things: whether people trust the answers, how good the answers are, whether people understand them and what they do next. Participants rate the answers and fill in monthly surveys. The study is due to run for more than 12 months, according to a report on its launch.

Apple is also frank about the risk. “Consumer health technologies powered by LLMs may at times offer responses that are incomplete or unclear,” its announcement says. LLMs, or large language models, are the technology behind chatbots such as ChatGPT.

The privacy terms are strict. Study data will not be used to train AI models. Apple says the data is made anonymous before it is sent to an outside cloud provider, and OpenEvidence deletes questions and answers after 24 hours, TechCrunch reported.

Most health chatbots skipped this step

More than 230 million people ask health and wellness questions on the most popular chatbot every week, its maker said in January. That figure is for ChatGPT alone. In July, OpenAI opened a separate health section in the app to users in the US aged 18 and over.

OpenAI has tested its models with a benchmark, a fixed set of questions marked like an exam. It grades answers to 5,000 realistic health conversations against rules written by 262 doctors. A test like this can show whether an answer is accurate. It cannot show whether a worried person understood the answer, or whether they called a doctor afterwards. That is the question Apple’s study is trying to answer.

Where the study falls short

The study has limits. Only people with a recent high-end iPhone can join, so the results may say little about people who find it hardest to see a doctor. The announcement also does not say whether or when Apple will publish what it finds.

Yet the app’s page in the App Store reads like an advert. The listing invites people to ask the app all their health questions and to build exercise or meal plans. It also says that neither the app nor the study gives medical advice, diagnosis or treatment.

Still, the order is right: test first, then sell. Last week we argued that people should not have to give up their privacy to use an AI assistant. Apple’s terms show that strict limits are possible: the data is not used to train AI, and questions are deleted after 24 hours.

Why it matters: Hundreds of millions of people already ask chatbots about their health each week, and there is little public evidence on whether those answers help them. A study this large could provide that evidence, but only if Apple publishes the results.

The post Apple Is Right to Test Its Health Chatbot Before a Full Launch appeared first on DailyAI.

LVMH Signs Deal to Show Articles From Its 17 News Brands in ChatGPT

Seventeen French media brands owned by LVMH, including Les Echos and Le Parisien, will have their articles summarised and linked in ChatGPT answers under a deal with OpenAI announced on Monday.

The agreement was signed on 16 November and also gives staff across the media group access to the paid business version of the chatbot. The group, called Groupe Les Echos-Le Parisien, has 1,800 employees, including nearly 700 journalists. It says its titles reach more than 27 million people in France every month. Its other brands include Paris Match, Challenges, L’Opinion, Investir and L’Agefi.

What changes for people who use ChatGPT

When someone asks a relevant question, ChatGPT will show summaries and short extracts from the group’s articles. Each answer will name the source and include a direct link to the original story. About 1.2 billion people use the chatbot each week, according to a separate announcement on 10 November.

The announcement does not give a price for the deal, and it does not say whether the articles will be used to train AI models. A similar deal signed in 2024 with a big American news group was reported to be worth $250 million. That group, News Corp, owns The Wall Street Journal.

What changes in the newsrooms

Teams will test AI for three jobs: researching documents, comparing documents and producing rough first translations. The stated aim is to give journalists more time for investigations, analysis and reporting in the field. Staff will get training as the tools are rolled out.

“Integrating AI will become an everyday asset for all our teams,” said Michèle Benbunan, chief executive of the media division. Emmanuel Marill, who runs OpenAI’s business in Europe, the Middle East and Africa, said OpenAI wants to help journalists “spend more time on what makes their journalism valuable” while respecting their editorial independence.

The group published the same announcement in French on its own website. Both versions say AI will also help build more personalised products for readers and advertisers.

Why it matters: More people now ask a chatbot for the news instead of visiting a news site. OpenAI said this month that ChatGPT runs about a fifth as many searches as Google. Outside estimates put that share much lower. For publishers, a deal like this pays off only if those links send readers back to their pages, and neither company has shared any figures on that.

The post LVMH Signs Deal to Show Articles From Its 17 News Brands in ChatGPT appeared first on DailyAI.

Amazon Caught AI Hacking Agents Using Stolen Keys to Fund More Attacks

For about a month, decoy servers run by Amazon recorded AI agents carrying out hacking attacks. The records show the agents stealing the keys that companies use to pay for AI models, and then using those stolen keys to power the search for more keys.

Amazon’s threat intelligence team published the findings on 13 November. From 3 October to 5 November, the decoy servers posed as AI services, and the attacking agents sent them their own notes on their campaigns. The agents ran on ARTEX, a free tool that was built for penetration testing, where experts attack a system to find weak spots before criminals do. Amazon says it has moderate confidence that the operators were separate groups linked to China. It points to prompts written in simplified Chinese and computers set to China’s time zone.

How the stolen keys paid for more attacks

An API key works like a password that lets software use a paid AI model. Whoever holds the key can use the model, and the key’s owner pays the bill. The agents looked for AI apps left open on the internet, took the keys stored inside, checked that they worked and added them to a shared pool. They went after tools that hold keys for many AI providers at once, such as LiteLLM, Dify and Ollama, because one weak admin page can expose many accounts.

Their playbook listed easy ways in: default admin passwords that nobody had changed, half-finished installs and known security flaws. The agents also searched public code, software packages, paste sites and Telegram channels for leaked keys. The people behind the campaigns told the agents to grade every key and to treat keys that had been public for more than 30 days as low value.

The records cover hundreds of targets. Amazon says it warned customers who may have been affected, but it could not check what the agents claimed to have done on systems it does not monitor.

Deepfakes failed to fool one identity check

In a second campaign, the agents tried to get past the identity checks at two cryptocurrency exchanges. They built a live deepfake video system on a single consumer graphics card that reached 11.5 frames per second. In their best attempt, the fake face moved to within 7 pixels of where the check asked it to be. Even so, the verification company’s server rejected the deepfake every time.

The same tool keeps turning up in other attacks. In a separate case this week, an attacker who used the tool with a stolen access token made 400 requests to copy databases in one hour. Earlier this month, another report found that companies were being charged more than $100,000 a day by criminals who had hijacked their AI access.

Why it matters: “None of the techniques in these campaigns were new,” Amazon’s team wrote. What changes is speed: AI agents never get tired of searching, so a key left in public code or an admin password left unchanged is now more likely to be found and used.

The post Amazon Caught AI Hacking Agents Using Stolen Keys to Fund More Attacks appeared first on DailyAI.

Top Candidate in South Korea’s Tax Exam Admits He Cheated With AI

The candidate who came first in South Korea’s tax accountant exam this year has admitted that he cheated with generative AI. Han Won-seok answered 159 of 160 questions correctly in the first round, a result that tax experts had called close to impossible.

Han, who is 34, confessed on Sunday in a video posted by his lawyer. He said he used a miniature camera to photograph the questions, sent the photos to an AI tool and read the answers on a smartphone. He said he used the same method in both rounds of the exam and in the first stage of this year’s exam for certified public accountants. He did not name the AI tool. He also refused to give more details, so that other people could not copy the method.

How his scores gave him away

Han scored 397.5 out of 400 in the first round in April. Three months later, he scored 663 out of 800 in the written second round and finished first overall. Four exam supervisors who watched him during the two rounds told the exam agency that they had seen nothing unusual. Doubts grew only after Han posted his scores on an online forum for tax accountants and claimed that he had passed through self-study.

The government’s plan to stop cheating with AI

On Monday, the finance ministry set up a task force with the National Tax Service and the Human Resources Development Service of Korea, which runs the exam. The tax service has started the process to cancel his result and to ban him from the exam for five years. The exam agency will file a criminal complaint for obstruction of business. Officials will also check the results of successful candidates from the past three to five years for unusual score patterns.

From next year, exam centres will have more security cameras, radio detectors that find hidden devices, better training for supervisors and a centre where people can report cheating.

The case comes days before South Korea’s national college entrance exam on Thursday. Ha Joon-kyung, the policy chief in the president’s office, held an emergency meeting. In that meeting, Mr Ha asked ministries to consider tools such as thermal imaging cameras. Last month, two college applicants were reported to have used AI glasses to try to cheat in university entrance exams.

Schools and universities face a similar problem. In a 2024 test at a British university, markers failed to spot exam answers written by AI. Teenagers in the United States also use AI for homework more often than a year ago.

Why it matters: Han was caught because his score was too high to ignore, not because the people watching the exam noticed anything. A candidate who used the same tools to get an ordinary pass mark might never be found, which is why officials will now recheck the results of the last three to five years.

The post Top Candidate in South Korea’s Tax Exam Admits He Cheated With AI appeared first on DailyAI.

Philippines Creates a National AI Council to Guide AI Policy

The Philippines now has a national council to guide how the government handles artificial intelligence. President Ferdinand Marcos Jr. created it with an executive order, at a time when AI is already changing work in an industry expected to employ about 1.94 million Filipinos this year.

Marcos signed Executive Order No. 131 on 12 November, and his press office announced it on Saturday. The full text of the order sets out what the new council must do. The order also adopts a national framework for AI governance as the guide for future policy. An executive order is an instruction from the president to government offices. It is not a law passed by Congress, and it does not add new rules for companies.

Members of the National AI Council

The National Artificial Intelligence Council reports directly to the president. The executive secretary chairs it. Its other members are the heads of ten government departments and agencies, including those for the economy, information technology, trade, finance, education and labour, plus one person from business. The president picks that business member for three years.

Five working groups will draw up plans for AI skills, data and research, computing infrastructure, investment and the institutions that oversee AI. Government agencies must update their progress on an online tracking system every quarter. The council must then report to the president within 30 days of the end of each quarter, and review the framework at least once every two years.

What the rules say about risk

The framework says government measures should match the level of risk an AI system carries. Possible tools include AI impact assessments, regulatory sandboxes, which let firms test products under supervision, and ways for people to file complaints. Existing regulators, such as those for banking or telecoms, keep their current powers.

The order also asks the council to propose ways to protect workers and students affected by AI. That matters in a country where call centres and other outsourcing firms are a major employer. The main industry group expects revenue to grow faster than jobs, and its chief executive has linked this to AI. The IT and Business Process Association of the Philippines forecasts $42.3 billion in revenue in 2026 and 1.94 million full-time workers, rising to 1.99 million in 2027.

Other governments in the region are moving faster on binding rules. Australia plans a dedicated AI safety regulator. Japan has drafted safety rules that would carry penalties.

Why it matters: The order puts one body in charge of AI policy in the Philippines, but it creates no new obligations for AI companies yet. The first test will be whether the working groups produce plans with real deadlines for the 1.94 million people who work in the outsourcing industry.

The post Philippines Creates a National AI Council to Guide AI Policy appeared first on DailyAI.