AI News Roundup – AI agents access user bank accounts, Major cybersecurity vulnerability discovered using AI models, AI and the US midterm elections, 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.

• AI agents have been detected accessing users’ bank accounts, in some instances without user instructions, according to Axios. Major U.S. financial institutions such as JPMorgan Chase, Bank of America, and Wells Fargo are actively working to contain the agents, which are entering banking platforms faster than security teams can adapt. While many customers willingly share their login credentials with consumer agents such as Meta’s Muse to automate financial tasks, this practice can violate most banks’ rules which often require direct human access. Further, banks often struggle to determine whether account activity is directed by a human customer or an autonomous bot that has malfunctioned. To combat the issue, several banks are moving to restrict external software to read-only visibility, similar to how some budgeting applications function. Other banks are responding by blocking third-party AI models from operating inside customer accounts and processing transactions. One banking industry source noted to Reuters that although institutions flag and block sessions involving shared credentials, transactions cannot technically be labeled as unauthorized if a customer knowingly provided their password. Still, leaders across Wall Street view such hazards as unacceptable. JPMorgan Chase CEO Jamie Dimon confirmed that his firm is developing internal tools to counter AI agent risks, while Charles Schwab CEO Richard Wurster explained that his company actively tracks and blocks outside agents to protect users’ accounts.

• The Wall Street Journal reports on a major cybersecurity vulnerability recently discovered with the assistance of generative AI tools. Known as PHP2Shell, the critical flaw affects PHP, a scripting language that powers up to 70% of websites globally. The issue stems from fundamental flaws in how PHP manages computer memory, enabling attackers to execute unauthorized code, compromise servers, and target website visitors. Although this software weakness was flagged by a researcher several decades ago, developers had long dismissed it as too difficult to exploit. However, researcher Ryan Austin at cybersecurity firm Calif utilized OpenAI’s GPT-5.6 Sol model during a recent source code audit to prove the vulnerability could be used to hijack a computer in as little as a few seconds on less complex systems. Volker Dusch, a security engineer at the foundation overseeing PHP, compared the severity of the flaw to the historic Heartbleed vulnerability, which affected hundreds of thousands of computer systems in 2014. Other security teams, including Google’s Wiz group, independently uncovered related exploits and confirmed that the vulnerability was likely too intricate for humans to detect without the use of AI. After PHP2Shell’s details were leaked prematurely on an online forum, software maintainers accelerated the release of a massive patch spanning nearly 200 files and 10,000 lines of code to prevent automated attacks, and cybersecurity experts recommend updating browsers and other internet-related systems as soon as possible.

• Less than two weeks away from the U.S. midterm elections, AI is having a major effect on how voters receive information online, according to The New York Times. Generative AI features and search overviews are now integrated across almost all major online platforms, presenting voters with automated summaries and synthetic content alongside traditional candidate information. A survey from the American Association of Political Consultants Foundation revealed that 20% of voters plan to use AI chatbots to research candidates, and twice that number have already encountered automated search summaries, even though only 1% view artificial intelligence as the “most credible” source for election information. Research organizations such as the Institute for Strategic Dialogue warn that human oversight has fallen behind the pace of AI development, leaving substantial gaps in safeguards against misinformation. U.S. lawmakers, including California Sens. Adam Schiff and Alex Padilla, have criticized companies such as Meta, TikTok, and xAI for providing incomplete answers regarding their AI safety protocols concerning election misinformation, though the companies have claimed that they are working on features to combat such issues in the lead-up to the elections in early November.

• Bloomberg reports on the release of Moonshot AI’s newest Kimi AI model. The Beijing-based startup announced Kimi K3.1 Pro as the first installment in a new series of AI models engineered for advanced computer programming, financial analysis, and 3D modeling for video game creation. This launch represents the company’s first major upgrade since its foundational K3 system debuted in July and achieved benchmark scores comparable to industry leaders OpenAI and Anthropic (as this AI Roundup covered at the time). The new model is positioned to maintain Moonshot’s competitive standing against domestic rivals like DeepSeek and Z.ai, whose recent releases had eclipsed Moonshot’s offerings in several major benchmarks. The release also follows a private financing round that established Moonshot’s valuation at approximately $50 billion ahead of an expected Hong Kong stock market debut early next year. The startup has rapidly expanded its commercial footprint through subscriptions for its Kimi Work and Kimi Code platforms, with annual recurring revenue exceeding $1 billion in August and approaching a $2 billion target for the year. However, Moonshot continues to encounter headwinds from U.S. officials and competitors who have alleged that Chinese developers train models by distilling outputs from American systems and thus violating their terms of service. The company may also be subject to pending Chinese regulations that mandate export registrations for deploying homegrown models abroad (as this AI Roundup covered last week).

• A new AI model allows biology researchers to “chat” with DNA sequences, according to WIRED. The tool, dubbed Carbon-X, was developed by open-source platform Hugging Face alongside researchers from Zhongguancun Academy in China and the University of Naples Federico II in Italy. The system combines natural language processing with genomic data, enabling both scientists and those without technical training to ask conversational questions about how genetic variations can influence gene function and human health. To power the interface, developers expanded a language model to read DNA in groups of six letters and trained it on 300 billion tokens across diverse organisms, including mammals, plants, and fungi, though the research team deliberately omitted viral sequences from the training dataset to guard against the creation of biological weapons. Carbon-X operates with 14 billion parameters, allowing the software to run locally on a standard laptop to protect sensitive patient records from being transmitted to external cloud servers, as many advanced AI models require. The model was also designed to analyze DNA segments of up to 50,000 base pairs to evaluate whether specific mutations are benign or may cause disease or other health issues. Despite the potential of such AI models, Carbon-X’s creators note that real-world lab testing is still required to confirm the accuracy of the model’s predictions.