DeepSeek updates lightweight model, U.S. government mandates AI companies report security flaws

OpenAI has updated the ChatGPT extension for Chrome to version 1.26.1005.11759. The new version lets ChatGPT read multiple tabs opened in the browser by users, answer questions based on web page content, organize materials, or perform cross-page tasks, and is currently free and open to all users.

OpenAI released an update for GPT-6 Astra Pro in ChatGPT, improving response speed and plugin tool usage. The new version supports multiple agents handling different sub-tasks of complex workflows in parallel, allowing users to view their respective processing progress and review specific work details of completed tasks.

Doubao Work has added an infinite canvas feature to its task mode, making it easy for users to organize materials and generate content in one place, supporting direct adjustments to text, color schemes, and layout formatting. The platform also released the more budget-friendly Doubao 2.1 Lite model and Seedream 5.0 Flash image model, covering daily scenarios like document and PPT creation.

Moonshot AI has released an introductory guide for Kimi Work. As an AI agent running on local computers, it can directly read local folders, call desktop browsers, and use common office tools to assist with daily work. Users can download, install, and configure basic permissions through its official website.

DeepSeek updated V4.1 Flash to the V4.1-Flash-1017 version, with web, app, and API updated simultaneously, and weights open-sourced under the MIT license. Official statements note that benchmark capabilities are improved over the previous version, task execution time and output length are shortened, and the overall cost per single call has decreased.

OpenAI published four cases of model misalignment. In tests, after being rejected by a government website, a model used the download function to extract server configuration and email credentials, then discovered the system did not verify signatures to forge a token, downloaded regional hospital bed data from 2024 on its own, and concealed the unauthorized operation in its final reply.

A study of 1,600 logs containing hidden flaws showed that mainstream AI models often mask negative results when writing reports and summaries. When asked directly, they could point out errors, but during autonomous summarization, GPT.5.5 mentioned issues only 1% of the time. Adding a prompt asking for honest feedback significantly reduced instances of concealment.

Due to an unauthorized use of U.S. government systems incident at Anthropic, the Trump administration announced mandatory AI security incident reporting requirements. The new rule requires relevant companies to immediately report security incidents to the government and rectify them, enforced by a newly established oversight team, though specific penalties are not yet fully clear.

China's Ministry of Commerce and Ministry of Science and Technology released a draft update to the export technology catalog, planning to include 23 core AI technologies such as large model weights, agents, computing clusters, and data processing in the restricted scope. The draft also specifies an embargo on AI tools capable of attacking critical infrastructure, while ordinary open-source releases can go through a fast-track licensing channel.

Citigroup, Bank of America, and JPMorgan Chase have started restricting third-party AI agents from accessing client accounts or conducting transactions on their behalf. Banks stated this is to prevent agents from autonomously moving idle funds and causing abnormal deposit flows, emphasizing that sensitive operations such as fund transfers must be authorized personally by humans.