AI News Roundup – OpenAI publishes more AI-discovered mathematical proofs, Open-weight AI models used to generate CSAM, AI-assisted hackers breach South Korea’s largest banks, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• The New York Times reports that OpenAI has published hundreds of further discoveries by its AI models in the field of pure mathematics. In a blog post, the company announced 377 new findings in varied mathematical fields, including algebra, number theory, logic, and topology. These new findings follow OpenAI’s much-heralded announcement that it found a solution to the Navier-Stokes problem, as this AI Roundup covered last month. Though, researchers recently found that the natural-language proof and the computer code proof generated by OpenAI do not match, which may cause issues with verifying the solution. One of the 377 findings released by OpenAI this past week is a purported proof of the “quasi-Riemann hypothesis,” which is related to the Riemann hypothesis, one of the Millennium Prize mathematical problems to which the Navier-Stokes problem also belongs. Many mathematics researchers are responding to these advances with consternation, with one professor saying that many mathematicians “felt like they got punched in the stomach.” Mathematics students in particular are worried about the future of their field and the value of their education, with one Ph.D. student telling the NYT that his colleagues have been asking “[w]hat can be done to ensure that a 2028 graduate has a degree with the same value of that of a 2023 graduate?” Computers have been used in formal mathematics for many years, though the scale and speed of the new findings have led some other mathematicians to question whether OpenAI understood their own research, with one professor telling the NYT that he “[didn’t] think they’ve done their sort of due diligence at all.” Indeed, since the 377 results were released earlier this week, OpenAI withdrew three papers and edited several others, with a spokesman saying that “[w]here errors are identified, we will work to correct them promptly and withdraw papers if no fixes can be found. As with other research manuscripts, this is an iterative process.”

• Popular open-weight AI models have been used by child predators to generate child sex abuse material, or CSAM, according to a feature from Bloomberg. Steven Anderegg, a Wisconsin software engineer, allegedly used the open-source AI image generator Stable Diffusion to generate sexually explicit images of children, which he later sent to a minor via Instagram. Anderegg was arrested in February 2024, and investigators found more than 13,000 such AI-generated images on his devices. Anderegg’s case is just one example of AI-generated CSAM fueled by open-weight AI models, which are freely available and can be tinkered with by those with the appropriate technical know-how. Investigators from around the world have targeted users of Stable Diffusion as well as the German Flux and Chinese Qwen and Hunyuan models. Since such models are run locally, user behavior is not tracked, allowing for CSAM generation to occur undetected. These materials have rapidly spread on the internet, often through the so-called “dark web,” and have grown in prevalence as the AI boom continues. Over 1.5 million reports of AI-linked CSAM were submitted to the U.S. National Center for Missing and Exploited Children in 2025, over 20 times the amount in the year before. Anderegg is still awaiting trial on federal CSAM distribution charges, though one of the original charges against him for CSAM possession was nullified by a U.S. appeals court in late August of this year, citing the U.S. Supreme Court’s decision in Ashcroft v. Free Speech Coalition, which held that possession of CSAM that does not depict real children was protected under the First Amendment’s freedom of speech rights. However, investigators say that AI-generated CSAM can still harm victims, as the AI models used in these instances are often trained on images of real child abuse, leading to real victims (or their approximations) appearing in generated material. In a statement to Bloomberg, Stability AI (the maker of Stable Diffusion) said that the company “is committed to preventing the misuse of AI technology, particularly in the creation and dissemination of harmful content, including CSAM,” and that they “have implemented robust safeguards to enhance our safety standards to protect our products from misuse,” though cybersecurity experts have found that conducting safety testing for AI models against CSAM generation is often difficult and can be legally risky for the researchers themselves.

• The Wall Street Journal reports that hackers used a Chinese-developed open-source AI agent to hack into the systems of South Korea’s largest banks and steal customer information. South Korean officials reported that the cyberattack hit at least seven major financial firms in the country, compromising the personal data, annual income details, and loan limits of roughly 68,000 customers and employees. The tool involved in the breach was an open-source cybersecurity agent called Artex AI, which was originally created by a Chinese engineer to help organizations find security weaknesses in their networks. The software works by drawing on different AI models (including models from Anthropic, OpenAI, and DeepSeek) to autonomously find security gaps in networks and plot attack paths with little human input needed. Investigators have not yet identified the specific group responsible for the breach, as the attacks came from more than two dozen internet addresses across roughly twelve countries, including the U.S., Japan, and Germany. In response to the incident, South Korean President Lee Jae Myung urged financial institutions to immediately upgrade their cyber defenses. Artex AI’s creator did not comment directly on the matter, but did update the tool’s user guidelines to prohibit the use of Artex AI for “malicious reasons.”

• Italy’s Prime Minister, Giorgia Meloni, has applied to trademark her voice to combat AI-generated deepfakes of herself, according to the Associated Press. Meloni submitted a trademark application with the European Union Intellectual Property Office including a four-second audio clip of herself saying “Io sono Giorgia Meloni” (“I am Giorgia Meloni”), echoing a Meloni campaign slogan from 2019 that later inspired a viral dance remix. As generative AI models have become more advanced, so-called deepfakes mimicking the likenesses and voices of humans have become increasingly lifelike and convincing, leading to major problems for public figures. Manipulated images of Meloni have circulated on social media sites in recent years, and the potential for such deepfakes to be used for malicious political purposes is also high. While Meloni’s trademark alone will not automatically prevent deepfakes, it would provide an additional avenue for her to combat deepfakes when they arise. Meloni’s lawsuit against a man who allegedly created sexually explicit deepfake images of Meloni remains pending in an Italian court.

• The Verge and Bloomberg report that Google and the game engine company Unity Software have announced an AI platform collaboration that will allow users to generate video games from text prompts. The platform, dubbed “Playground,” will “put game creation in the hands of millions of new creators, and bring new experiences to billions of players worldwide,” according to a joint statement from the companies. Playground is based on Google’s past work on foundational world models (which this AI Roundup has covered in the past) which generate entire virtual environments as opposed to text, images, audio, and/or video like most generative AI models. The companies also announced Unity Spark, an AI-powered game engine, for more advanced game creation based on Unity’s existing offerings for game developers. Playground was released this past week, while the companies stated that Unity Spark will enter a beta testing phase soon.