EU plans to expand Europol's powers and double its budget, intensifying privacy oversight disputes(original, opens in a new tab)

The European Commission has proposed a transformative reform of Europol, doubling its budget to €3 billion and having the agency, which has coordinated European police intelligence since 1998, directly engage in countering new threats and gain access to more data and advanced technology. European Data Protection Supervisor (EDPS) head Wojciech Wiewiórowski expressed support for Europol's use of data and AI tools, but questioned the lack of accompanying oversight arrangements in the proposal and warned that the agency could retain large amounts of data on people unrelated to criminal activity for long, unspecified periods. Europol Deputy Executive Director Jürgen Ebner said AI makes crime faster, more complex, and easier to scale, and that AI can help investigators filter massive volumes of digital evidence, connect information, and identify actionable intelligence. The proposal retains the EDPS's supervisory role over Europol, but Wiewiórowski believes the EDPS's powers and resources have not been strengthened in tandem, and that internal bodies have too much power without external oversight.

promptfoo 0.124.0 released: adds support for Claude Opus 5.5, Sonnet 5.5, GPT-6 Sol/Luna, and Grok 4.7(original, opens in a new tab)

promptfoo released 0.124.0, adding support for models such as Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Sol and Luna, GPT-6.1 Sol, and Grok 4.7, and adding OpenAI safety identifiers. This release contains breaking changes: the hosted ChatKit provider was removed, and SDKs such as WatsonX, Langfuse, Slack, Codex Security, and local Transformers are now installed on demand. It also fixes issues with assertion scoring, CLI exit codes, and Cloud login, and upgrades the MCP SDK to 1.32.0 to fix a redirect security issue.

ChatGPT generates fake The New Yorker cartoons and forges cartoonists' signatures(original, opens in a new tab)

A Nieman Lab investigation found that ChatGPT-generated cartoons in The New Yorker style carry the pen-name signatures of real cartoonists, involving more than 15 current and former contributors to the magazine, including Brendan Loper, Emily Flake, Joe Dator, and Pat Byrnes, none of whom were authorized or compensated. Loper's signature was first flagged to him by a stranger in May as appearing in ChatGPT output, and afterward a large number of imitation cartoons with real signatures appeared on Reddit's ChatGPT forum. An OpenAI spokesperson said the company values unexpected model behavior reported by the community, and after being informed by journalists, ChatGPT began prompting that the prompt may violate third-party content similarity guardrails, but as of press time it still signed some generic cartoons with real cartoonists' names. Condé Nast signed a multi-year content licensing agreement with OpenAI in 2024, but The New Yorker said it has never allowed any large model developer to train models on its cartoons; the magazine's cartoonists work under freelance submission contracts, with authors retaining copyright.

AI safety groups question OpenAI's reduction in political spending(original, opens in a new tab)

POLITICO reports that AI safety groups do not endorse OpenAI's posture of reducing political spending. Leading the Future, supported by OpenAI President Brockman and his wife, as well as early investors Andreessen and Horowitz, has invested $75 million; meanwhile, Guardrails Alliance, founded in June, has raised $5 million, with some of the funds coming from OpenAI employees who oppose Leading the Future. Some Democratic candidates supported by Leading the Future have declined to state whether they will cut financial ties with the organization.

Appeals court grants xAI temporary relief on Minnesota's AI undressing law(original, opens in a new tab)

CBS Minnesota reported that the Eighth Circuit approved xAI's injunction pending appeal in a single sentence, without explaining its reasoning. This cannot be taken to mean the court has ruled the law unconstitutional, or that all companies' related obligations have been eliminated. The report covers regulation of AI-generated sexualized images and victim relief.

Why it was selected明尼苏达去衣法被上诉法院暂停执行，呈现了生成式 AI 内容监管与第一修正案之间的司法拉锯。

Florida woman used Claude as a diary to write attack threats; Anthropic alerted police after human review(original, opens in a new tab)

Carli Michelle Heller of Bonita Springs, Florida, was arrested for writing content in Claude planning an attack on a sheriff's office, facing a second-degree felony charge of written threats of violence under Florida law. She wrote the content on September 26 and said she used Anthropic's chatbot as a diary. Claude's safety system flagged the entry and escalated it to human reviewers, who determined it constituted a credible threat and reported it to law enforcement. Anthropic said it may share user information in limited emergency situations when it believes disclosure is necessary to prevent death or serious bodily harm. Officers used this to confirm Heller's identity and went to her home, detaining her without incident.

Anthropic's Cowork moves model inference and VM to the cloud(original, opens in a new tab)

Anthropic's Felix Rieseberg announced that the new version of Cowork runs both model inference and the VM in the cloud, with each session having an isolated sandbox that does not share state with other sessions. The old version ran an Anthropic-provided VM locally to execute tool calls, and users were unhappy with its disk, battery, and performance overhead, and work would be interrupted when closing the laptop. In the new version, when the VM needs to access files on the user's device, the desktop app handles the file access tool calls.

Meloni applies to EU for voice trademark to counter AI deepfakes(original, opens in a new tab)

Italian Prime Minister Giorgia Meloni has filed a sound trademark application with the European Union Intellectual Property Office, using two four-second voice samples of "Io sono Giorgia Meloni," and the application is currently under examination. This move comes ahead of Italy's general election next year, as a large number of deepfake videos impersonating her to promote investment schemes have recently appeared on social media, and her voice clone has also been used in campaign ads urging voters to oust her. Lawyer Guido Scorza believes that filing a trademark infringement complaint makes it easier to get platforms to remove content than filing a personal data complaint; Gianni Riotta, head of the Italian Digital Media Observatory, noted that trademark protection is unlikely to curb the scale of deepfake spread, and its real value lies in raising public awareness.

Sony Music requests removal of over 260,000 AI deepfake songs(original, opens in a new tab)

Sony Music is intensifying its fight against streaming fraud, having requested platforms to remove over 260,000 tracks by the end of September, nearly double the 135,000 at the end of March. Sony says generative AI deepfake songs that imitate its artists' voices and images without permission are appearing on a large scale, involving current and past artists such as Harry Styles, Adele, Britney Spears, Queen, and Michael Jackson. Beyond deepfakes, streaming platforms are also flooded with AI-generated tracks mass-produced for repeated bot plays. Dennis Kooker, President of Global Digital Business at Sony Music, said fraudulent plays could account for 10% of tracks on streaming platforms, and record label executives estimate they cause up to $2.2 billion in losses to the industry annually. Deezer said its platform receives about 90,000 AI-generated songs uploaded daily, accounting for more than half of new music, and that up to 85% of plays of fully AI-generated tracks last year were fraudulent.

CLTC 与 SaferAI 联合发布 AI 风险建模开放问题论文(original, opens in a new tab)

CLTC 的 AI Security Initiative（AISI）与欧洲非营利机构 SaferAI 于 2026 年 3 月举办线上研讨会，聚焦 AI 安全中研究不足的风险建模领域，22 名来自民间社会、学术界和政府机构的专家参会。会后 12 名参会者发布论文《Open Problems in AI Risk Modeling》，梳理了限制传统方法应用于风险建模的方法论与制度挑战，并比较了基于场景的风险估计与基于贝叶斯网络的阈值设定两种主流方案。论文指出目前尚无单一成熟的 AI 风险建模方法，并提出涵盖模型结构、范围、证据整合、验证与治理的开放问题议程。

Anthropic 安全裁决对政府承包商意味着什么(original, opens in a new tab)

针对 Anthropic 的一项安全裁决将影响政府承包商。Blank Rome LLP 在 JD Supra 上分析了该裁决对政府承包商的影响。

U.S. Senator Schatz pushes AI safety regulation legislation, vote blocked by Ted Cruz(original, opens in a new tab)

U.S. Senator Brian Schatz of Hawaii, along with Democratic colleagues Mark Warner and Andy Kim, attempted to push a bill to a vote that would establish an AI Safety Committee under the U.S. Department of Commerce and require new models to be submitted to the federal government before public release. The vote was ultimately blocked by Republican Senator Ted Cruz, who said the warnings were scary but he was unwilling to give such power to the executive branch, and advocated a cautious bipartisan approach to address threats to critical infrastructure, nuclear, and biosecurity. In an interview with Civil Beat, Schatz said AI risk is urgent, that he cannot judge whether the probability of catastrophe is 10% or 2%, but he is unwilling to tolerate catastrophic risk at that level; he mentioned the Hugging Face incident as a turning point in public understanding of the risk, and advocated introducing third-party independent evaluation from the development stage rather than testing only before release.

vLLM fixes denial of service vulnerability where overly loose cache_salt validation allowed EngineCore to be taken down by a single request(original, opens in a new tab)

On deployments where vLLM enables the LMCache-MP KV connector, cache_salt validation is insufficient. A special format can trigger an exception in downstream cache key parsing, and the scheduling path does not isolate it as a single request error, causing the shared EngineCore to exit and interrupt concurrent services. The impact is availability under specific connector configurations, with no data leakage or code execution involved. The GitHub advisory marks affected versions as < 0.30.0 and the fixed version as 0.30.0.

Why it was selectedThe announcement gives the complete call chain from request fields to EngineCore crash and two fix recommendations, and teams running vLLM inference services can use this to investigate their own deployments.

vLLM structured output vulnerability can crash the engine; fixed in 0.30.0(original, opens in a new tab)

Three structured output paths in vLLM can let request errors escape to the shared EngineCore, including syntax compilation exceptions, negative tokens from speculative decoding, and empty structured constraints accepted by the optional Rust frontend. An ordinary API request can terminate the engine, interrupting concurrent and subsequent service for all tenants on that instance. The impact is limited to availability, and does not involve code execution, memory corruption, or data leakage. The GitHub advisory marks affected versions as < 0.30.0, with the fix in 0.30.0.

Why it was selectedThe announcement gives the specific code locations and fix patches for three structured output paths that caused EngineCore to go completely down, and teams deploying vLLM can use this to troubleshoot versions and configurations.

vLLM multimodal transmission trusts caller fields, 0.30.0 fixes crash and cache poisoning vulnerabilities(original, opens in a new tab)

vLLM's scale-out multimodal transport treats caller-supplied features as trusted rendering results without binding validation against the current model's rendering contract. Five trust boundary flaws can let authenticated requests cause shared engine crashes, cross-request encoder cache poisoning or disclosure, and loss of transport integrity. Report submitted by Patch the Planet (a Trail of Bits and OpenAI collaboration), discovered using GPT-5.5-Cyber. The GitHub advisory marks affected versions as < 0.30.0 and the fixed version as 0.30.0.

Why it was selectedThe announcement itemizes the trigger points and impact of five forged fields, along with a fix patch; teams deploying vLLM multimodal inference can use this to investigate similar trust boundaries.

vLLM discloses a GLMGA video sampling vulnerability that can cause request-driven CPU and memory exhaustion(original, opens in a new tab)

vLLM's GLMGA video sampler constructs an index list based on request-controlled sampling parameters, then deduplicates and reads the video. Small videos combined with overly large parameters can also consume disproportionate CPU time and memory, delaying other media requests and causing partial denial of service. The advisory does not report memory corruption, data leakage, or code execution; video options can be filtered at the gateway, and untrusted clients can be restricted from selecting this sampler. The GitHub advisory marks affected versions as >= 0.23.0rc2, < 0.30.0, with the fixed version being 0.30.0.

Why it was selectedThe announcement gives the cause and fix recommendations for the denial-of-service in the vLLM video sampling interface, and teams deploying multimodal inference services can use this to check request-level parameter boundaries.

vLLM's Qwen2-VL/Qwen3-VL video sampler does not enforce the request-level max_frames limit, which can lead to unauthenticated memory exhaustion(original, opens in a new tab)

vLLM's Qwen2-VL and Qwen3-VL video samplers do not enforce the existing num_frames cap on request-level max_frames, which can lead to memory exhaustion. In the advisory's CPU-only test, decoded frames increased from 60 to 900 and peak RSS increased from 2,271 MiB to 13,629 MiB; these values correspond only to that test. Paths such as /tokenize do not require vLLM API-key authentication; the Rust frontend is unaffected because it rejects the relevant media options. The GitHub advisory marks affected versions as >= 0.24.0, < 0.30.0 and the fixed version as 0.30.0.

Why it was selectedThe report provides measured data on unauthenticated requests exhausting vLLM memory and the affected version range; teams deploying Qwen-VL services can use this to check versions and request parameters.

vLLM tool continuation loses cache_salt, restoring cross-tenant prefix cache membership inference(original, opens in a new tab)

In the /v1/responses tool continuation path of GPT-OSS Harmony, vLLM loses cache_salt, allowing continuations to enter a global unsalted prefix cache. An attacker can use cache counters to infer whether another tenant has processed a guessed prompt or history, breaking tenant isolation. This vulnerability was reported by Patch the Planet (a collaboration between Trail of Bits and OpenAI) and discovered using GPT-5.5-Cyber. The GitHub advisory marks affected versions as < 0.30.0 and the fixed version as 0.30.0.

Why it was selectedIt discloses the specific code location where vLLM loses cache_salt on the Harmony tool continuation path, which teams deploying multi-tenant inference services can use to troubleshoot prefix cache isolation.

Self-modeling interventions can modulate emergent misalignment in models(original, opens in a new tab)

Tagade et al. study how self-recognition and self-reporting modulate emergent misalignment, finding that some training interventions can prevent or reverse some metrics, while self-reported data can also propagate misalignment. The authors acknowledge that the operationalization is somewhat coarse, some experiments use only a single seed, and the agentic evaluation is not connected to real tools, so they cannot claim general recovery.

研究：九款对话式 AI 服务向第三方追踪器泄露对话标题、提示词与永久链接(original, opens in a new tab)

IMDEA Networks 等机构的研究者对 ChatGPT、Claude、Gemini、Grok、DeepSeek、Perplexity、Copilot、Mistral 和 Meta AI 九款对话式 AI 服务的网页客户端及其中八款的安卓客户端做了系统隐私分析，共识别出 44 个第三方组织、124 个第三方域名，每款服务都至少接入一个广告或追踪服务。研究发现 6 个网页客户端和 3 个安卓客户端会把对话 URL、标题、提示词和截图传给第三方，其中 5 个网页客户端向 9 个追踪服务泄露对话永久链接或全局对话 ID，3 个网页客户端泄露对话标题，且常与持久用户标识同时传输。部分提供商的对话永久链接默认公开可读，Grok 免费与付费层默认公开并可选择退出，Perplexity 访客层对话始终公开。

Why it was selected研究覆盖九款对话式 AI 的网页与安卓客户端，量化了对话标题、提示词和永久链接流向第三方追踪服务的规模。

AI email summarizers can be deceived without hidden text(original, opens in a new tab)

A test of AI email summarizers showed that a payload containing no hidden text and no explicit instructions to the summarizer produced fabricated meeting dates and invoice amounts in the summary in all 10 trials. The researchers used 6 emails to combine the two variables of hidden text and instructions, ran each 10 times for a total of 60 trials, and added 10 clean emails at each end as controls, registering the true and fabricated facts and the judging criteria in advance. The results showed that fabricated information was written into the summary in all 10 trials, while only emails that directly instructed the summarizer reliably removed true facts; a sample padded with just 30 blank lines retained only two true facts across ten trials at temperature 0. The authors concluded that both hidden-text detection and instruction-keyword detection have gaps; instruction detection can block some attacks, but cannot prevent fabricated information from being swallowed whole by the summarizer. The test was conducted only in one email client and one viewport, without verifying performance at higher temperatures or in other clients.

研究笔记：从预训练数据中过滤绕过监督相关信息(original, opens in a new tab)

Geodesic 与 Redwood 的作者介绍了从零预训练 30B 模型时过滤绕过监督相关知识的初步研究。作者称过滤后知识题表现下降，而一般能力保持。作者明确表示尚未测试真实的绕控行为是否因此减少。

Microsoft UFO's Mobile MCP press_key tool can inject ADB commands(original, opens in a new tab)

The press_key tool in Microsoft UFO's Mobile MCP server has a command injection vulnerability that allows authenticated callers to execute arbitrary shell commands on an authorized Android device. The key_code parameter of press_key is a free-form string that is appended to adb shell input keyevent without validation, and the adb client joins the arguments after shell with spaces into a remote command without escaping, so shell metacharacters in key_code enter the Android remote shell command. Exploiting this vulnerability requires a valid UFO_MCP_API_KEY bearer token, an accessible Mobile MCP server, and adb installed on the host with an authorized device connected. The advisory states that this does not by itself constitute host OS command execution, Android root execution, or access to private app data beyond the Android shell user's permissions.

San Diego County sues AppLovin, Roblox and Polymaker, alleging circumvention of parental controls and child safety risks(original, opens in a new tab)

圣迭戈县新成立的消费者公平与公共保护部门宣布对 AppLovin、Roblox 和 Polymaker 提起三起诉讼，涉及儿童安全、消费者隐私和 3D 打印幽灵枪。针对数字广告平台 AppLovin 的诉讼称，即便家长开启了保护控制，其技术仍绕过这些防护，在儿童使用的手机游戏中投放成人交友、酒精、电子烟、大麻、露骨色情和暴力广告，并收集可能暴露儿童住址、学校及睡眠状态等敏感设备信息。针对 Roblox 的诉讼称其向家长和儿童宣传平台是安全的游玩与社交场所，却未采取足够措施阻止成年人冒充儿童接触未成年用户。针对 3D 打印材料商 Polymaker 的诉讼称其明知产品可用于制作无序列号的幽灵枪仍进行宣传，并推出枪械主题产品、与枪械打印网红合作、付费推广。县政府依据加州虚假广告法和不正当竞争法提出指控，寻求禁令、赔偿和民事罚款。

Why it was selected圣迭戈县新设消费者保护机构的首批诉讼，展示了地方层面对广告平台绕过家长控制、收集儿童数据的具体指控路径。

FoMo-FD：用流匹配世界模型监测手术机器人模仿策略的执行失败(original, opens in a new tab)

Zhefeng Huang 等人提出 FoMo-FD，一种面向手术机器人模仿策略的执行级运行期失败监测器，只使用成功执行数据训练，无需收集失败示范。它把多视角相机画面经冻结的 DINOv2 与特征空间 β-VAE 压缩为视觉隐变量，用流匹配世界模型预测短时窗口的标称视觉-动作动态，再通过逆传输把实际端点隐变量回推到基高斯空间，以非符合性得分与符合预测校准的阈值比较来触发告警。在 Needle Pickup、Ring over Post 仿真任务和 dVRK 上的 Tissue Retraction、Shunt Insertion 任务中，覆盖 20 种失败模式、320 次失败与 80 次成功 rollout，腕部相机监测达到 96.6% 失败检出率与 1.3% 误报率，固定相机为 45.9% 与 10.0%。

巴黎公立医院 AP-HP 报案：Facebook 上 AI 克隆寄生虫学教授推销排虫产品(original, opens in a new tab)

巴黎公立医院集团 AP-HP 与索邦大学于 10 月 5 日通报，Facebook 上出现三段用 AI 伪造 Pitié-Salpêtrière 医院寄生虫学专家 Renaud Piarroux 教授形象与声音的视频，宣传一款号称能清除体内寄生虫的产品，未经其本人同意。AP-HP 已报案并联系 Meta 要求下架，相关视频被删除后，Facebook 上又出现新的类似视频。AP-HP 指出，这类内容不仅损害医生和医院形象，还可能误导公众使用缺乏科学数据支持的产品，危害患者健康。此前 2 月，营养学家、Nutri-score 创始人 Serge Hercberg 医生的形象也曾被同类视频冒用。

数学与 AI 顾问组就 OpenAI 发布数学成果发表意见并重申发布准则(original, opens in a new tab)

数学与人工智能顾问组（AGMAI）就 OpenAI 发布由内部模型生成的大批数学成果发表意见，称 OpenAI 报告了数百个开放问题的解答，并认为这是数学界的重要事件，其影响远超单个结果本身。该顾问组由 François Charles、Timothy Gowers、Martin Hairer、Edward Witten 等数学家组成，在 OpenAI 接触部分成员提议设立外部顾问委员会后，改为独立组建，不接受任何 AI 公司报酬，也不具备决策权。顾问组此前于 9 月 29 日发布 AI 生成数学成果负责任发布的一般准则，参考了六百多份问卷回复；10 月 6 日的意见重申这些准则，表示已与 OpenAI 讨论并认可其沟通意愿，但强调评估准则落实程度仍应由数学界完成。

安特卫普大学警告 Deepfake 视频冒用疫苗学家 Pierre Van Damme 推销前列腺药物，Meta 起初拒绝下架(original, opens in a new tab)

安特卫普大学警告社交媒体上流传一段 Deepfake 视频，冒用该校疫苗学家兼流行病学家 Pierre Van Damme 的名义推销一款所谓前列腺问题特效药，视频中还出现 VRT 主持人 Xavier Taveirne 在演播室宣布该药上市。Van Damme 声明自己从未录制该视频，其面容、声音和姓名均未经同意被使用。校方联系 Meta 后，Meta 以内容不明确违法为由两次拒绝处理，第三次交涉后仅在一个页面删除该视频，相同或类似视频很可能仍在其他页面流传，Van Damme 与校方正在考虑进一步法律行动。

英国律师纪律庭因 AI 幻觉引文将一名律师除名(original, opens in a new tab)

英国律师纪律庭（SDT）因律师在法庭文件中使用 AI 生成的虚假判例引文而将其除名，Kennedys 律所对该案作出解读。该律师在事后纠错时再次借助 AI，又生成了不存在的判例，最终被除名。这是英国律师纪律庭首次因 AI 幻觉引文对律师作出处分。

Microsoft FARA 7B success rate on WebVoyager drops from 74% to about 38%(original, opens in a new tab)

Microsoft researcher Corby Rosset and Browserbase's Miguel González Fernández point out that Microsoft's FARA 7B web agent achieves a 74% success rate on the WebVoyager benchmark as scored by the official judge, but drops to about 38% when evaluated with a general-purpose verifier calibrated against human annotations, with a Cohen's Kappa of 0.58 against human labels. The article's analysis notes that existing LLM judges exhibit systematic flaws in long-trajectory evaluation, such as ignoring scoring rubrics and multimodal context overflow, and that evaluation reliability needs to be improved through rubric alignment and process scoring.

Why it was selected同一模型同一基准，官方 GPT-4o 判分给出 74%，经人工校准的通用验证器只有约 38%，差距来自判分器设计而非模型本身。

Joshi v. OpenAI Foundation case: court issues order on 33 motions to dismiss(original, opens in a new tab)

U.S. District Judge Mark E. Walker of the Northern District of Florida issued an order on the 33 motions to dismiss in Joshi v. OpenAI Foundation (4:26-cv-00222), setting a consolidated briefing schedule. Defendants may file a consolidated motion to dismiss and memorandum of no more than 12,000 words by October 21, 2026, with plaintiffs' response due November 4.

俄罗斯运营商 Beeline 用 AI「赛博家庭」陪聊诈骗电话：今年转接 1910 万通、拖住骗子 14 万小时(original, opens in a new tab)

俄罗斯运营商 Beeline 将疑似诈骗电话转接给模拟奶奶、青少年、母亲和父亲声音的语音 AI 助手，让其与骗子长时间闲聊。自今年年初以来共有 1910 万通电话被转给这套「赛博家庭」，AI 助手累计与诈骗分子通话约 14 万小时（近 6000 天）。同期 Beeline 反欺诈系统拦截超 8.7 亿通骚扰电话，积累的对话被用于再训练模型，以更有效地识别并拖住骗子。

Colorado woman falsely accused of theft after Flock license plate hit, sues Flock Safety and two town police departments(original, opens in a new tab)

Colorado woman Chrisanna Elser has filed a class-action lawsuit against Flock Safety, the towns of Columbine Valley and Bow Mar, and two police officers over being falsely accused of theft in September 2025. The complaint alleges that officer Jamie Milliman identified her as the thief of a $25 package and issued a citation based solely on Flock's license plate recognition system showing her truck passing through Bow Mar during the incident, and refused to review vehicle camera and GPS records that could prove her innocence. Elser spent two weeks clearing her name before the police chief dismissed the citation. The complaint also alleges that the camera network Flock sold to the two towns covers all three entrances and exits of Bow Mar, allowing officers to search the historical movements of any vehicle without a warrant, probable cause, or supervisor approval, and cites patent filings and audit logs claiming the system has facial recognition, predictive policing, and demographic analysis capabilities, and has long lacked controls such as search justification and multi-factor authentication.

Why it was selectedThe indictment details how the Flock license plate recognition system was searched by police without a warrant or reasonable suspicion, leading to a Colorado woman being falsely accused of theft, useful for assessing compliance risks in ALPR deployment.

安特卫普大学警告 Deepfake 视频：病毒学家 Pierre Van Damme 被伪造代言前列腺药物(original, opens in a new tab)

安特卫普大学警告一段 Deepfake 视频，视频中病毒学家 Pierre Van Damme 被伪造代言一款前列腺问题药物，他本人强调从未录制该视频，其面部、声音和姓名均未经同意被使用。视频由主持人 Xavier Taveirne 在虚假电视演播室中引出，Van Damme 的伪造形象随后推销产品约 2 分钟，已有观众询问是否应停用现有治疗。校方向 Meta 举报后，前两次均被以内容不违法为由拒绝删除，第三次举报后视频才从 1 个页面移除，但疑似仍有相同或类似视频在其他页面流传，Van Damme 与校方正在考虑进一步法律行动。

vLLM Semantic Router / KR Labs releases Vela 2.0 open-source routing and guardrail model family(original, opens in a new tab)

vLLM Semantic Router and KR Labs released the Vela 2.0 open-source model family, with four sizes—0.3B, 0.8B, 4B, and 9B—for routing, safety checks, and span-level decisions, with weights and code released under Apache-2.0. The models receive options, labels, and scoring criteria at request time and return structured answers with probabilities and character offsets; a single call can handle routing, safety, PII, and hallucination issues simultaneously. The official statement says the 9B model achieved 0.989 AUC on prompt attack detection for unseen attack families, higher than Vela 1.0 Guard's 0.792; multilingual hate speech reached 0.855 AUC, higher than 0.646. The 0.3B model can run on CPU, with a single full routing request taking 0.09 seconds on an A40, while the 9B takes 0.60 to 0.71 seconds.

巴黎公立医院 AP-HP 报案：AI 深度伪造寄生虫学教授代言排虫产品(original, opens in a new tab)

巴黎公立医院集团（AP-HP）就 Facebook 上利用 AI 克隆其医生形象和声音推销排虫产品的虚假广告报案。AP-HP 称，三段视频未经同意使用了 Pitié Salpêtrière 医院寄生虫学专家、索邦大学教授 Renaud Piarroux 的形象，宣传一款号称能清除体内寄生虫的产品，实为 AI 合成的伪造内容。AP-HP 已向 Meta 投诉并成功使三段视频下架，但此后仍有类似视频不断出现。AP-HP 同时警告健康风险，指出这类内容可能让网民相信虚假信息，或使用缺乏科学数据支持有效性与安全性的产品，并强调该机构不会在社交网络上为任何产品或服务做商业推广。

俄罗斯运营商 Beeline 用 AI「赛博家庭」陪聊诈骗电话：今年转接 1910 万通、拖住骗子 14 万小时(original, opens in a new tab)

俄罗斯运营商 Beeline 将疑似诈骗电话转接给由 AI 语音助手组成的「赛博家庭」陪聊，用奶奶、少年、妈妈、爸爸等声音拖住骗子。公司总经理 Сергей Анохин 称，自今年年初以来已有 1910 万通电话被转给「赛博家庭」，累计通话时长 14 万小时，约合 6000 天；同期反欺诈平台还拦截了超过 8.7 亿通骚扰电话。这些对话还被用作模型再训练素材，使 AI 助手能力不断增强。

研究评估嵌入预过滤，将百万级 LLM Agent 社交网络消息筛查成本降低最高 65.6%(original, opens in a new tab)

研究者提出用轻量嵌入预过滤器先筛掉明显安全的内容，再交给前沿 LLM 判定，以降低对 Agent 社交网络 Moltbook 全量消息做安全筛查的成本。团队用 GPT-5.5 以高推理强度标注了 1 万条 Moltbook 帖子与评论，其中 9.2% 被判定为严重度 3 及以上的不安全内容；两名人类标注者之间一致性较高（κ=0.823），但与 LLM 判定仅为中等一致（κ=0.578）。在召回率校准到 0.80 时，扫描 787,226 条消息的预估成本从 7085 美元分别下降 49.9%（MiniLM-L12-v2）、55.6%（BGE-M3）和 65.6%（有监督分类器），其中 MiniLM 的推理速度约为分类器的 32 倍。数据集 MoltSafe-10K 与复现代码已公开。

比利时教授 Pierre Van Damme 遭 Deepfake 冒名推广"神药"，有患者考虑停掉癌症治疗(original, opens in a new tab)

一段伪造比利时教授 Pierre Van Damme 的 Deepfake 视频在网上推广所谓"治疗前列腺炎的灵丹妙药"，Van Damme 本人回应称"我很不好受"。报道称有民众因此考虑中断自己的癌症治疗。

Indian court in Chandigarh orders Facebook to suspend three pages over AI-faked political videos(original, opens in a new tab)

印度昌迪加尔法院于 10 月 5 日下令 Facebook 暂停三个页面，理由是这些页面发布了针对旁遮普国大党人士 Amandeep Kumar 的 AI 生成视频。法院认为相关内容超出正常公共批评范围，可能损害民主与公共共识的形成，并指出视频带有政治倾向、偏向旁遮普执政党。涉案页面为 Batale di Rajneeti、Kachi laan 和 Thanedar uncle，分别约有 2200、1100 和 11000 名关注者，相关视频播放量约 17.8 万、18 万和 4.78 万次。法院批准临时禁令，要求页面在下次开庭前停止传播涉及 Kumar 的 AI 生成内容，并指示平台暂停这三个页面的公开访问，案件将于 10 月 22 日再次开庭。