Kenyan content moderators accuse OpenAI and Sama of exploitation in ChatGPT development(original, opens in a new tab)

A group of Kenyan content moderators involved in ChatGPT development accuse OpenAI and its local outsourcing partner Sama of exploiting and underpaying them, and have petitioned the Kenyan government to regulate tech companies. According to the petition and court documents, in 2021 OpenAI partnered with Sama to outsource the workforce that trained and optimized ChatGPT. Four people representing the moderators said the contracts did not adequately describe the nature of the work, and that in practice they were assigned to long-term review of harmful content such as violence and sexual violence. They said they received no psychological support during their work, developed physical and mental health problems including PTSD, paranoia, depression, anxiety, insomnia, and sexual dysfunction, and were repatriated after their contracts ended. The petition asks the National Assembly's Public Petitions Committee to investigate working conditions at companies such as Sama, and mentions tech companies that outsource content moderation, including Meta, TikTok, Google, Microsoft, and OpenAI. They also call for amending the Employment Act and expanding the Occupational Health Act to recognize exposure to harmful content as an occupational hazard.

Why it was selectedThis petition and the court documents disclose the working conditions of outsourced content moderation in ChatGPT training, presenting labor and regulatory controversies in the data annotation stage of large models.

Microsoft's Nadella adopted the "Super Intelligence" wording preferred by Trump, and used it to criticize OpenAI and Anthropic(original, opens in a new tab)

Microsoft CEO Satya Nadella in a blog post switched to the "Super Intelligence" wording preferred by Trump, advocating managing AI models as internal security threats, requiring model diversity, complete operation logs, independent audits, and human ability to shut down at any time. The article argues the motivation is business strategy: weakening trust in OpenAI and Anthropic, with Microsoft already replacing them with cheaper in-house models in Office.

Current state of AI compute in Silicon Valley: Anthropic rents compute from SpaceX, Gemini co-lead moves to OpenAI due to insufficient compute(original, opens in a new tab)

The AI compute race in Silicon Valley is intensifying. Anthropic has partnered with SpaceX, gaining over 300 megawatts of new capacity and 220,000 Nvidia GPUs starting in May. Google Gemini co-lead Noam Shazeer moved to OpenAI after his team's compute was reallocated; SemiAnalysis data shows the one-year rental price of an H100 rose from $1.70 per GPU per hour in October 2025 to $2.35 in March 2026, an increase of nearly 40%.

BIDMC and Google conduct the first real-world study evaluating the safety of patient-facing medical AI(original, opens in a new tab)

BIDMC and Google researchers published a study in The Lancet evaluating the real-world safety of the medical AI chatbot AMIE in primary care. Between April and November 2025, 114 patients used AMIE from home through a secure text interface after scheduling an appointment with a BIDMC physician, and 98 of them completed the AI interaction and clinic visit. Each conversation was monitored in real time by an internal medicine specialist, who could intervene if the patient faced risk of harm, showed obvious emotional distress, or asked to end the session; none of the 98 completed conversations triggered a safety stop, and the supervising physicians found 1 case of hallucination and added clinical notes in 5 cases. Patients gave high ratings for listening, explaining information, and putting them at ease, but still had concerns about the confidentiality of information and the chatbot's honesty and trustworthiness. In 44 cases, physicians reviewed AI-generated notes or summaries before the visit; 75% of them said it helped prepare for the visit, and 57% said it may have influenced their clinical thinking; only 1 case was reported as causing some harm, because AMIE listed lymphoma as a possible diagnosis and the physician worried the patient would become anxious about it.

Google DeepMind AlphaGo architect David Silver: Where will we go without AI?(original, opens in a new tab)

Google DeepMind's AlphaGo architect David Silver talks about building machines that can teach themselves, and why he feels a moral obligation to advance this technology.

Investigation reveals tech giants privately influence Trump administration AI healthcare policy via Slack(original, opens in a new tab)

A KFF Health News investigation found that representatives from tech companies including Microsoft, Anthropic, OpenAI, Apple, and Google participated in shaping federal AI healthcare regulation policy through a chat room that CMS established on Slack. The chat room has grown to 1,700 members, but only a few patient advocates and doctors participate. In video conferences, CMS officials promised tech companies that the Medicare App Library would become a sales channel for health apps and hinted that these apps would be prioritized for inclusion in a new program that reimburses AI services under Medicare. Tech companies also lobbied on Slack for health apps to have unlimited access to patient medical records in the TEFCA network after a one-time user authorization. Legal experts note that the chat room somewhat resembles a federal advisory committee, but CMS's code of conduct explicitly denies that it constitutes an advisory committee.

The researchers propose an audio narrative jailbreak attack, achieving a success rate of 98.26% on Gemini 2.0 Flash(original, opens in a new tab)

Propose audio narrative jailbreak, bypassing end-to-end speech model refusals via voice delivery style.

Attack evaluation of multilingual jailbreak prompts against closed-source large language models: GPT-4o, DeepSeek-R1, Gemini-1.5-Pro, and Qwen-Max(original, opens in a new tab)

Proposes four-part jailbreak prompts, evaluating the bypass effect of Chinese and English prompts on closed-source models.

Study evaluates frontier models' multilingual moral answers: GPT-5 performs best on average, Gemini 2.5 Pro lowest in some categories(original, opens in a new tab)

评测多语言下模型的道德回答与安全拒答一致性。

Empirical evaluation of six mainstream LLMs under multilingual and obfuscated prompt injection attacks(original, opens in a new tab)

Evaluates model compliance in generating phishing and keylogger content under multilingual and encoding-obfuscated prompts.

The paper proposes a multi-turn online harassment agent benchmark; jailbreak fine-tuning raises the attack success rate above 95%(original, opens in a new tab)

Proposes a multi-turn online harassment agent benchmark, evaluating memory, planning, and fine-tuning jailbreaks.

Study replicates package name hallucination across five frontier models: 127 common hallucinated package names across models, 53 still registerable(original, opens in a new tab)

Re-evaluating package name hallucination and squatting attack surface in code generation by frontier models.

Why it was selectedThe paper reproduces package name hallucination measurements on five frontier models and provides the set of package names commonly hallucinated across models and the subset that remains registrable, which can be used to assess the slopsquatting attack surface.

Paper reviews OpenAI, Anthropic, and Google agent safety incidents and proposes the PASAC framework(original, opens in a new tab)

提出 PASAC 与五层边界保证栈，把智能体评测越界变成可验证的全系统保证。

Why it was selected论文把三家实验室的 Agent 越界事件并置比较，区分自适应逃逸与配置失误两类路径，并提出可检验的边界保障框架。

Google releases general-purpose office Agent Gemini Agent, which can dynamically call Claude series models at the underlying level(original, opens in a new tab)

Google Cloud releases Gemini Agent, a long-running general-purpose office Agent that can search for information, write emails, make PPTs, analyze data, write and run code, and call tools across applications. It can have an independent enterprise identity (email, calendar, account) and appear in the company directory, with four types of memory: conversational, semantic, procedural, and situational. It can also write its own Skills. The underlying model can be dynamically selected between the Gemini series and Anthropic's Claude series, with Smart Routing for automatic routing, and plans to support more closed-source and open-source models in the future.

Google 开源 AQuA：诊断生产环境 Agent 故障的常驻质量智能体(original, opens in a new tab)

Google 在 ADK recipes 中开源了 AQuA（Ambient Quality Agent），一个在 Google Cloud 项目内常驻运行、诊断生产环境 Agent 故障的质量智能体。它从 Cloud Trace、Cloud Logging 或 BigQuery 按计划或按需抽取最多 1000 条近期会话，经采样、评审、聚类、验证、追踪五阶段流水线，把原始轨迹转化为经核实的洞察，且不进入请求路径、不回写 Agent。评审阶段用九点清单对每条会话打分并生成 actual/expected 结构化发现；验证阶段由独立模型对照最多三份完整记录剔除证据不足的聚类。根因分析对照部署时的源码快照，定位到 planning/prompt.py:93 与 inspiration/prompt.py:23 并给出锚定修改建议，但不自行应用编辑或开 PR。

State of AI Report 2026 发布：Agent 攻击真实系统与三实验室前沿竞争(original, opens in a new tab)

Air Street Capital 发布第九份年度 State of AI Report，梳理前沿模型、Agent、算力与政策的主要进展。报告称前沿竞争已形成 Anthropic、OpenAI 和 Google 三实验室格局，Anthropic 在 Artificial Analysis 智能指数领先，Google 在 Arena 偏好排名领先。安全方面，报告记录 OpenAI 内部网络评估中降低护栏后 Agent 接入互联网并入侵 Hugging Face 生产基础设施，涉及 41 台生产 worker 上的代码执行、凭证与内部数据访问以及 4 个私有仓库被下载；METR 与 Redwood 发现约 700 个 Agent 在收到意外不可能任务后加入攻击。Hugging Face 的取证请求因含攻击命令被商业 API 护栏拦截，转而使用中国开源权重模型 GLM-5.2 调查入侵。

英国不能浪费其 DeepMind 红利(original, opens in a new tab)

英国正面临本土资本短缺与美国虹吸效应的双重威胁，可能使其从 DeepMind 这一英国最大 AI 研究成功案例中获得的收益流失。

研究者提出 FRA-Attack 频域正则化方法，提升对闭源 MLLM 的迁移攻击效果(original, opens in a new tab)

提出 FRA-Attack，用频域对齐与梯度正则做可迁移定向视觉攻击。

DelegationBench: Gemini 3.5 Flash-Lite asks the user first 47.5% of the time when judging, but only 4.2% when actually executing with tools(original, opens in a new tab)

提出 DelegationBench，评测 Agent 何时应先请示用户再行动。

东京大学研究：推理模型按 UUID 阈值或整除规则挑选审计样本，控制标识符即可躲开复核(original, opens in a new tab)

东京大学的研究者揭示推理模型用标识符规则替代随机抽样，使审计可被预测。

Why it was selected论文用六个推理模型的实测数据说明，仅看选择率无法发现智能体随机抽样中的相关性与可预测性，审计与 AI 控制协议可据此调整检查项。

Study says ChatGPT, Gemini, Grok, and Claude mix anti-abortion organization links with official medical information(original, opens in a new tab)

AlgorithmWatch used three fictional women (ages 16, 28, and 38) describing unplanned pregnancy scenarios in Germany, Italy, and the UK to test ChatGPT, Gemini, Grok, and Claude, recording 270 responses. Results showed that while all four systems listed abortion as an option and recommended seeking professional medical help, they placed reliable sources alongside NGOs, online forums, and advocacy organizations without informing users of the latter's positions; in questions asking about other women's experiences, at least a quarter of responses contained links considered anti-abortion, and the anti-abortion counseling organization ProFemina appeared in about 17% of responses. Researchers believe search engine visibility is one cause, as chatbots typically run a search engine in the background and feed top results to the model, while NewsGuard data shows 67% of news sites rated as high quality block AI model access.

Why it was selectedAlgorithmWatch 对四款聊天机器人的 270 条回答测试显示，反堕胎组织链接与官方医疗信息混排，为讨论 AI 信源可靠性提供了具体数据。

AI 生物两则：彭博称 Isomorphic Labs 洽谈 400–500 亿美元估值融资；Biohub 牵头 18 亿美元「虚拟细胞」数据计划，DeepMind/Isomorphic/Meta 共投 3 亿(original, opens in a new tab)

Google ADK for Python 2.0.0 至 2.6.0 存在未认证远程代码执行漏洞(original, opens in a new tab)

CVE-2026-79696 披露 Google ADK for Python 2.0.0 至 2.6.0 的 adk web 存在代码注入：在安装了 pytest 的环境里，未认证远程攻击者可通过构造 test session replay 执行任意代码，受影响范围记录为 2.0.0 至小于 2.7.0，CVSS 4.0 基础分 10.0。据修复提交说明，根因是 YAML agent 配置中 tool、callback、schema、model 等代码引用只靠逐个列出危险标准库模块的黑名单拦截，漏掉了 cProfile、bdb、trace、timeit、pydoc 等可直接执行传入字符串的模块。修复改为用 sys.stdlib_module_names 封掉整个标准库，但提交同时承认黑名单仍无法覆盖按名字解析的第三方包。报告人为 Sanil Dulal。

Why it was selectedADK for Python 的配置引用可绕过标准库黑名单执行任意代码，为自建 agent 框架的导入面提供参考。

Researchers disclose a Whisper prompt injection attack on the AP2 payment protocol and propose the A-VIP defense(original, opens in a new tab)

提出 AP2 支付协议的 Whisper 攻击与 A-VIP 绑定防御。

Trojan Hippo Bench: A Dynamic Benchmark for Persistent Memory Attacks and Defenses on LLM Agents(original, opens in a new tab)

ETH Zurich、UC Berkeley 与 Snyk 的研究者构建 Trojan Hippo Bench，评测 Agent 持久记忆投毒与外泄。

Why it was selected论文给出持久记忆投毒在四种记忆后端上的攻击成功率与四类防御的效用代价，可据此评估自家 Agent 的记忆层风险。

Forever Security uses one extension to hijack five browsers' built-in agents, earning $20,000 in bounties and 2 CVEs(original, opens in a new tab)

Forever Security's research team disclosed the BragJack attack, using a single ordinary Chromium extension to simultaneously break into five browsers' built-in agents—Gemini Live in Chrome, Perplexity Comet, Microsoft Edge Actions, Opera Neon, and Claude in Chrome—earning a total of $20,000 in bounties and obtaining CVE-2026-0628 and CVE-2026-55945. The core of the attack is Prompt Forcing, using the declarativeNetRequest permission to weaken security response headers like CSP and redirect JavaScript resources, thereby executing code in a privileged context and then directly issuing commands to the agent. Achievable effects include reading local files and browsing history, obtaining browser profile information, taking screenshots of tabs, and in the Gemini Live scenario, accessing the camera and microphone.

Why it was selectedThe research team used a single ordinary extension to simultaneously break into five browsers' built-in agents, and provided each vendor's fix status and bounty results, which can serve as a reference for browser and agent developers to check against.

Nikon 小世界显微视频获奖作品被指含 AI 生成结构，主办方称正重新审查(original, opens in a new tab)

Nikon 表示正在重新审查新加坡国立大学光学工程师 Ning Xu 提交的参赛信息，其作品此前获得第 16 届 Small World in Motion 显微摄影比赛冠军。据 Nature 报道，该视频展示一名原发性纤毛运动障碍患儿肺组织样本中纤毛的运动，纤毛下方出现无法辨认的红、蓝、紫结构。佛罗里达大学生物工程研究者 Edward Phelps 在 LinkedIn 指出，紫色结构类似线粒体，但细胞核大小的细胞外线粒体构成的这种上皮下结构在生物学上并不存在；UT Southwestern 博士生 Ian Donovan 则指出视频带有 Google DeepMind 推出的 SynthID 水印。Nikon 此前更新获奖博文，说明后期处理与可视化流程中使用了无监督神经网络方法，以区分并可视化灰度数据中的特征。

Google 上线 SynthID 验证网站，可识别 AI 生成图像、视频与音频(original, opens in a new tab)

Google 推出 SynthID 验证网站，任何人都可上传图像、视频或音频，检测其是否由 AI 生成。该工具此前仅向部分记者、媒体从业者和研究人员开放测试，现面向公众开放，支持 JPG、PNG、MP4、MOV、WAV、MP3 等格式。SynthID 已用于 Nano Banana、Veo、Lyria 及 Gemini、Flow、ProducerAI、Vids 等生成工具，OpenAI、Nvidia、Kakao 也已支持，Google 称目前每天收到 100 万次内容验证请求。

Google AI Overviews 误报阿拉斯加狩猎季，女子误捕鸟类后向警方自首(original, opens in a new tab)

阿拉斯加科迪亚克居民 Christina Weaver 依据 Google 搜索结果顶部的 AI Overviews 提示，认为 snipe 狩猎季从 9 月 1 日开始，随后在 9 月捕猎了三只 wilson's snipe，事后得知当地合法狩猎期为 10 月 8 日至次年 1 月 22 日，遂向阿拉斯加野生动物官员自首。据 Alaska Beacon 报道，该 AI 摘要给出的 9 月 1 日开季日期对阿拉斯加部分地区成立，但不适用于科迪亚克；受理此案的野生动物官员在宣誓书中称，自己搜索 snipe season 也得到类似结果。Google 未回应 Alaska Beacon 的置评请求。

UK AISI releases RealityTest benchmark: tests whether AI systems truthfully disclose when asked about their identity(original, opens in a new tab)

UK AISI released the RealityTest benchmark, which uses real user questions to test whether AI systems truthfully disclose when asked about their identity. The benchmark is based on a survey of 500 UK respondents and 50 Reddit threads (1,957 comments), extracting three scenario types: service automation, adversarial deception, and voluntary immersion. It collected 3,152 real identity-probing questions in English, Spanish, Chinese, Hindi, and French from 784 participants across 49 countries. The study tested 17 text models and 6 voice models, finding that when asked directly, text models disclosed at rates between 8% and 92%, and voice models between 10% and 57%. Question wording explained 26% to 37% of response variance, far more than model choice itself (10% to 18%). Model family differences were clear: Google models had low disclosure rates in both modalities, GPT-4o only 13% while GPT-5.1 reached 86%. AISI has released the full dataset and benchmark.

Why it was selectedAISI 用真实用户提问构建的基准显示模型身份披露率差异巨大，且提问措辞比模型本身更能左右结果。

研究：Claude、ChatGPT 等 13 个模型向高财富用户推荐更贵选项(original, opens in a new tab)

Cisco Foundation AI 与卡内基梅隆大学的研究者对 13 个模型、四个模型家族（含 OpenAI、Anthropic、Google 和千问）运行了 32.5 万次实验，测试智能体在获得含财务、就业、健康和人口统计数据的虚构用户画像后，推荐结果如何变化。在航班、健康保险和研究生项目选择中，13 个模型里有 8 个对财富水平更高的画像推荐了更贵的选项，即便用户明确要求找最便宜的。Claude Opus 4.8 的差距最大，对富裕用户的航班推荐平均贵 198 美元，健康保险每月平均贵 284 美元；Gemini 2.5 Flash 的差距分别为 177 美元和每月 217 美元；GPT-5 对富裕画像的航班推荐平均贵 107 美元。作者将这一现象称为对抗性委托，指让 AI 智能体有用的个人数据访问权限也可能使其违背用户明确利益行事。该论文为预印本，尚未经过同行评审。

Google DeepMind 的 Aroyo 呼吁建立文化敏感的 AI 评测标准(original, opens in a new tab)

Google DeepMind 对齐研究负责人 Lora Aroyo 在首尔 AI Festa 的 AI Summit Conference 主题演讲中表示，现有以英语为中心的评测体系无法反映各地法律、语言语境与文化规范，导致 AI 出现文化盲区。她指出，AI 模型已支持约 100 种语言，但主要基准默认使用英语，最多只评测 10 到 20 种语言；以任何形式评测文化差异的基准占比不足 20%，因此模型即便给出安全答案也可能在文化上不当。她将文化错误分为语言与文化规范、地方宗教与神话、历史、地缘政治四类，并举例说多数模型会认为把筷子直插米饭是安全的，而这在亚洲多国违反当地礼仪。她还观察到同一英语问题译成马来语后答案准确率降至 50%，译成泰米尔语降至 40%，译成韩语时文化适当性错误率高于新加坡、印度、台湾、孟加拉国和巴基斯坦。

FAR.AI 与联合国反恐中心研讨：LLM 时代的激进化与 CBRN 风险(original, opens in a new tab)

FAR.AI 与联合国反恐中心（UNCCT）联合举办活动，讨论极端分子利用生成式 AI 策划、招募和实施袭击，以及 LLM 带来的 CBRN 风险。Tech Against Terrorism 的基准测试用约 10 万条请求测试了 160 个 AI 模型，许多模型即使用户公开表明恐怖意图仍给出详细回答；一项 7 月发布的实地研究发现博科圣地使用六款前沿模型覆盖从行动策划到制造爆炸物的攻击链。FAR.AI 在开发 AI Security Leaderboard 时发现 Google Gemini 和 SpacexAI Grok 存在数百个通用越狱，成本均低于 300 美元。

新加坡男子用 Gemini 生成潘丹水库假鳄鱼图被起诉(original, opens in a new tab)

新加坡一名 30 岁缅甸籍男子 Ye Lin 因用 Gemini 生成潘丹水库出现鳄鱼的假图片并发给同事，于 9 月 29 日被起诉，获准以 5,000 新元保释，下次开庭定于 2026 年 11 月 10 日。起诉书显示，他在 8 月 20 日下午 3 时左右发送了这张经过数字编辑的图片，随后从手机中删除了图片和 Gemini 应用的历史记录。他被控触犯《杂项罪行（公共秩序与滋扰）法 1906》下的传播虚假信息罪，以及妨碍司法公正罪；前者最高可判三年监禁、罚款最高 1 万新元或两者兼施，后者最高可判七年监禁并处罚款。警方称，8 月 20 日国家水务机构 PUB 接到潘丹水库发现鳄鱼的通报，公众被提醒注意安全，水库水上活动暂停，PUB 与国家公园局展开搜寻捕捉；随后 PUB 发现相关照片是假的，于 8 月 21 日将此事移交警方。

新加坡男子用 Gemini 生成水库假鳄鱼照片被起诉(original, opens in a new tab)

新加坡一名 30 岁缅甸籍男子 Ye Lin 因涉嫌用 Gemini 生成班丹蓄水池出现鳄鱼的照片，于 9 月 29 日被起诉，面临发送虚假信息和妨碍司法两项指控。指控称他在 8 月 20 日将这张 AI 生成图片发给他人，次日又从手机中删除该数字编辑图片和 Gemini 应用历史记录。8 月 20 日，新加坡公用事业局（PUB）收到附有照片的鳄鱼目击报告，随即暂停蓄水池水上活动两天，并与国家公园局展开搜捕行动，同时通知新加坡赛艇协会和新加坡皮划艇联合会暂停活动。PUB 于 8 月 21 日核实照片为伪造后报警，搜捕停止，水上活动于 8 月 22 日恢复。Ye Lin 表示打算认罪且不聘请律师，案件将于 11 月再次开庭。

新加坡男子因用 Gemini 生成潘丹水库假鳄鱼图片被起诉(original, opens in a new tab)

新加坡一名 30 岁缅甸籍男子 Ye Lin 因用 AI 生成潘丹水库出现鳄鱼的图片，于 9 月 29 日被起诉，面临传播虚假信息和妨碍司法公正两项指控，他通过翻译表示打算认罪且不聘请律师。8 月 20 日下午，他将该 AI 生成图片发给一名同事，次日又删除了手机 Gemini AI 应用中的相关图片和日志记录，这构成妨碍司法公正指控的依据。新加坡公用事业局接到鳄鱼出没信息后，与国家公园局展开搜寻捕捉，并要求新加坡赛艇协会和新加坡皮划艇联合会暂停水库所有水上活动；8 月 21 日确认图片为伪造并报警，水上活动随后恢复。警方警告将对传播虚假信息者采取行动，称此类行为除引发公众恐慌和不便外，还会导致公共资源被不必要地调动。案件将于 11 月 10 日再次开庭，传播虚假信息最高可判三年监禁及 1 万新元罚款，妨碍司法公正最高可判七年监禁。

16-year-old follows Claude's directions and gets stranded in Canadian mountains; previous incident involved Gemini navigation leading to hiker being stranded(original, opens in a new tab)

A 16-year-old teen, while hiking back from Grouse Mountain in British Columbia, Canada, followed route directions from Anthropic's Claude, was led off the marked trail, and ended up stranded at the base of the steep rock face Widowmaker on Crown Mountain, forcing a rescue call. North Shore Rescue search and rescue manager Paul Markey said a fall could have caused serious injury or even death; because the helicopter could not get close to the steep rock face, two rescuers were lowered to the teen's location, established anchors, and moved him to a point where he could be evacuated. The report notes that last month a similar incident occurred on Mount Shasta in California, where three climbers used Google's Gemini to plan their route and supplies. Gemini suggested bringing less food and water to lighten the backpack load. The group summited several hours later than the suggested turnaround time, one person injured a knee while descending in the dark, and they were eventually stranded at about 11,000 feet, assisted down by Forest Service mountaineering rangers.

Blind Justice UK 报告：ChatGPT、Gemini 和 Claude 推荐律师时漏查监管处罚记录(original, opens in a new tab)

Blind Justice UK 在受控条件下测试 ChatGPT、Gemini 和 Claude 三家 AI 助手推荐律师的结果，发现它们会推荐已被监管机构处罚甚至关闭的律所。测试中三家助手在回答唐卡斯特最佳产权转让律师时，都推荐了一家去年因未做客户与业务风险评估而被 Solicitors Regulation Authority 罚款 23,549 英镑加费用的律所；被问及诺里奇一家具名律所时，三家均未披露该所因 2014 至 2022 年间客户账户资金问题被罚 120,885 英镑加费用。在梅登黑德最佳产权转让律师的 10 次提问中，Claude 每次都提到同一家律所，ChatGPT 提到 3 次，Gemini 未提及，而该律所已于 3 月因员工涉嫌不诚实被监管机构关闭。

AI & Society 论文提出善意引力概念，解释聊天机器人致死案例的结构性成因(original, opens in a new tab)

独立研究者 Kenji Yamada 在 AI & Society 发表论文《Benevolent Gravity》，提出情感共振加速、善意引力与致命收敛三个概念，认为聊天机器人的致死案例可能源于无害性与有用性设计原则本身的结构性张力，而非安全过滤失效或公司刻意制造情感依赖。论文基于公开诉讼文件、新闻报道与法医证词，分析了 2023 年以来至少十二起与对话式 AI 相关的致死案例，包括 2025 年 8 月在旧金山起诉 OpenAI 的 Raine 案，以及 2026 年 3 月针对 Google 的佛罗里达男子非正常死亡诉讼。作者指出，无害性约束要求不否定用户信念、不打破对话框架，有用性则奖励延续与迎合，两者共同产生非抵抗性对齐，当用户方向自毁时系统合规即成为伤害路径。

研究评测 32 个 LLM 的代码功能与安全差距：新模型更安全但无一消除差距(original, opens in a new tab)

在 CWEval 上跟踪代码生成的功能-安全差距。