十部门联合印发AI伦理审查试行办法，全面规范人工智能合规与安全

简报：

工业和信息化部等十个部门联合印发《人工智能科技伦理审查与服务办法（试行）》，旨在规范和促进人工智能科技活动的伦理治理，确保合规与安全。
办法明确了AI伦理审查的适用范围、实施主体及工作程序，涵盖申请受理、审查评估与专家复核等环节。
重点审查领域包括人机融合系统、具备舆论引导能力的算法模型，以及具有高度自主能力的决策系统。
提出了标准建设、服务体系、创新鼓励、宣传教育和人才培养等五个方面的支持举措，帮助企业提升风险防控能力。
对违反相关规定的行为明确了法律后果，将依据相关法律法规进行调查处理。

相关链接：

工信部发布《人工智能科技伦理审查与服务办法（试行）》明确治理新规

Ten Departments Jointly Issue Trial Measures for AI Ethics Review, Comprehensively Regulating AI Compliance and Safety 🤖

Briefing:

Ten departments, including the Ministry of Industry and Information Technology, have jointly issued the Measures for the Ethical Review and Services of Artificial Intelligence Science and Technology Activities (Trial), aiming to regulate and promote the ethical governance of AI-related scientific and technological activities and ensure compliance and safety.
The Measures clarify the scope of application, implementing entities, and working procedures for AI ethics review, covering application acceptance, review and assessment, and expert re-evaluation.
Key areas of review include human-machine fusion systems, algorithmic models with the capacity to guide public opinion, and decision-making systems with a high degree of autonomy.
Five supporting measures are proposed, covering standards development, service systems, innovation incentives, public awareness and education, and talent cultivation, to help enterprises strengthen their risk prevention and control capabilities 📘.
Legal consequences for violations of the relevant provisions are clearly specified, and such cases will be investigated and handled in accordance with applicable laws and regulations ⚖️.

DeepSeek V4开启密集灰测，全新多模态架构与视觉能力曝光

简报：

4月7日消息，DeepSeek V4正在进行密集灰度测试，在底层架构、交互逻辑和多模态能力上均有重大升级。
曝光界面显示其推出三大核心架构：主打响应速度的快速版（Lite）、基于全新记忆架构的专家版，以及可直接处理图视频任务的视觉版（Vision）。
DeepSeek深度适配本土算力生态，正开发至少2款完全基于国产AI芯片的大模型，测试申请优先对国内厂商开放以协同优化。
市场期待V4延续超长文本处理优势（预计支持百万级Token），并猜测或将推出专门针对代码生成的AI编程特别版。

相关链接：

DeepSeek V4 灰度测试曝光：全新视觉版现身

请问是否还需要为您总结其他相关的人工智能行业资讯？

DeepSeek V4 Enters Intensive Gray-Scale Testing, Revealing a New Multimodal Architecture and Visual Capabilities 🚀👀

Brief:

On April 7, reports indicated that DeepSeek V4 is undergoing intensive gray-scale testing, with major upgrades across its underlying architecture, interaction logic, and multimodal capabilities.
The leaked interface shows that it introduces three core architectures: a Lite version focused on response speed, an Expert version built on a brand-new memory architecture, and a Vision version capable of directly handling image and video tasks.
DeepSeek is deeply adapting to the domestic computing ecosystem and is developing at least two large models built entirely on Chinese-made AI chips. Testing access is being prioritized for domestic vendors to support coordinated optimization.
The market expects V4 to continue its advantage in ultra-long text processing, with support anticipated to reach the million-token level, and there is also speculation that a special AI coding edition tailored specifically for code generation may be launched. 💡

谷歌AI概览准确率约90%，仍面临高频错误与虚假信息挑战

简报：

据《纽约时报》报道，谷歌AI概览（AI Overviews）准确率约为90%，在庞大的搜索基数下，每小时可能生成超5700万条错误答案。
初创公司Oumi评估显示，Gemini 3的准确率较前代提升至91%，但其概览内容与原始信息来源不符的比例却升至56%。
测试表明AI概览极易受虚假信息影响，存在引用人工炮制的虚假博客以及在同一搜索页面出现自相矛盾信息的情况。
针对上述评估结果，谷歌发言人提出质疑，认为Oumi的测试方法未能真实反映实际用户的搜索行为。

相关链接：

谷歌搜索 AI 概览准确率仅 90%，易受虚假信息影响

Google AI Overviews Are About 90% Accurate, but Still Face Frequent Errors and Misinformation Challenges 🤖⚠️

Brief:

According to The New York Times, Google AI Overviews are about 90% accurate, but given the massive scale of search traffic, they could still generate more than 57 million incorrect answers per hour.
An evaluation by startup Oumi found that Gemini 3’s accuracy improved to 91% compared with its predecessor, but the share of overview content that did not match the original source information rose to 56%.
Tests showed that AI Overviews are highly vulnerable to misinformation, including cases where they cited fabricated blog posts and displayed contradictory information on the same search results page.
In response to these findings, a Google spokesperson questioned the evaluation results, arguing that Oumi’s testing methods did not accurately reflect real user search behavior.

Anthropic发布顶级网络安全模型Mythos，精准挖掘数千陈年代码漏洞

简报：

Anthropic正式推出新一代前沿模型Mythos的预览版，该模型在智能体级别的编程与推理及网络安全领域表现卓越。
在近期封闭测试中，Mythos成功扫描出数千个零日漏洞，并精准锁定了隐藏在一二十年前老旧代码中的高危历史遗留问题。
因具备极强的攻击性潜力，Mythos暂不对公众开放，目前仅通过“Project Glasswing”安全项目向苹果、微软等行业巨头及美国联邦官员定向试用。
因拒绝将技术用于对美国公民的自主打击或监控，Anthropic近期被美国国防部列为“供应链风险实体”。

相关链接：

Anthropic推出网络安全AI模型Mythos，专治代码陈年漏洞

还有其他资讯需要我为您整理成简报吗？

Anthropic Unveils Top-Tier Cybersecurity Model Mythos, Accurately Unearthing Thousands of Long-Standing Code Vulnerabilities 🔐

Brief:

Anthropic has officially launched a preview of its next-generation frontier model, Mythos, which delivers outstanding performance in agent-level programming, reasoning, and cybersecurity.
In recent closed testing, Mythos successfully scanned and identified thousands of zero-day vulnerabilities, while precisely uncovering high-risk legacy issues hidden in outdated code written one or even two decades ago.
Due to its exceptionally strong offensive potential, Mythos is not yet available to the public. It is currently being selectively trialed only through the “Project Glasswing” security initiative by industry giants such as Apple and Microsoft, as well as U.S. federal officials.
After refusing to allow its technology to be used for autonomous strikes or surveillance against U.S. citizens, Anthropic was recently designated by the U.S. Department of Defense as a “supply chain risk entity.” ⚠️

智谱GLM-5.1发布：编程基准夺冠并全线提价，国产模型迈向性能溢价

简报：

智谱正式发布全新大模型GLM-5.1，在衡量软件开发能力的SWE-bench Pro基准测试中刷新世界纪录，超越Claude 4.6 Opus位居全球首位。
该模型具备极强的自主工程能力，支持单次任务持续工作达8小时，可独立完成规划、执行、测试及错误自修复，直接交付工程级成果。
伴随新品上线，智谱GLM全线模型提价10%，核心场景价格已锚定Claude 3.5 Sonnet，标志着国产大模型竞争重心从成本倾销转向性能溢价。
受发布消息影响，智谱股价在4月8日上午一度大涨14.12%，公司总市值攀升至3964亿元。

相关链接：

智谱发布旗舰模型GLM-5.1：SWE-bench评分全球领跑

Zhipu GLM-5.1 Released: Tops Coding Benchmarks and Raises Prices Across the Line, as Chinese Models Move Toward Performance Premiums 🚀

Brief:

Zhipu has officially launched its new large model, GLM-5.1. In the SWE-bench Pro benchmark, which measures software development capabilities, it set a new world record, surpassing Claude 4.6 Opus to rank No. 1 globally.
The model demonstrates exceptionally strong autonomous engineering capabilities, supporting continuous work on a single task for up to 8 hours. It can independently handle planning, execution, testing, and error self-repair, directly delivering engineering-grade results.
Alongside the new release, Zhipu has raised prices by 10% across the entire GLM model lineup. Pricing for core use cases is now benchmarked against Claude 3.5 Sonnet, signaling that competition among Chinese large models is shifting from cost-driven underpricing to performance-based premiums 📈
Driven by the release announcement, Zhipu’s stock surged as much as 14.12% on the morning of April 8, pushing the company’s total market capitalization to RMB 396.4 billion.

目录

十部门联合印发AI伦理审查试行办法，全面规范人工智能合规与安全

Ten Departments Jointly Issue Trial Measures for AI Ethics Review, Comprehensively Regulating AI Compliance and Safety 🤖

DeepSeek V4开启密集灰测，全新多模态架构与视觉能力曝光

DeepSeek V4 Enters Intensive Gray-Scale Testing, Revealing a New Multimodal Architecture and Visual Capabilities 🚀👀

谷歌AI概览准确率约90%，仍面临高频错误与虚假信息挑战

Google AI Overviews Are About 90% Accurate, but Still Face Frequent Errors and Misinformation Challenges 🤖⚠️

Anthropic发布顶级网络安全模型Mythos，精准挖掘数千陈年代码漏洞

Anthropic Unveils Top-Tier Cybersecurity Model Mythos, Accurately Unearthing Thousands of Long-Standing Code Vulnerabilities 🔐

智谱GLM-5.1发布：编程基准夺冠并全线提价，国产模型迈向性能溢价

Zhipu GLM-5.1 Released: Tops Coding Benchmarks and Raises Prices Across the Line, as Chinese Models Move Toward Performance Premiums 🚀