Tag
文章介绍了 OpenAI 为满足欧盟 AI 法案第 50(2) 条来源透明度要求而推出的文本水印 textGrain，并展示其发布的失效曲线：同义词替换 10% 时检测率从 92% 降至 66%，替换 25% 时仅剩 17%。作者还构建了一个交互式攻击实验室，可实时演示翻译回译、同义词替换和数学化文本如何让水印检测的 p 值崩塌，强调未检测到水印并不能证明内容出自人类之手。
该论文提出一种将LLM代理的自进化限制在运行时harness（指令文本、工具调用逻辑与原语组合）的框架，模型权重保持固定，使每次适应都成为可审查的diff，并通过哈希链记录的准入门进行管控。模拟实验显示受控双循环仅放行144/7449个候选变更且未恶化误差，而无约束系统放行了309个有害变更。
This paper extends the Auditable Agents framework's Policy Checkability dimension to claims about AI agents, defining when a claim is auditable and proving that independent record coverage, per-action authorization binding, and completeness are necessary for verification, with practical guidance for operators and standard setters.
OpenAI will enable automatic watermarking of ChatGPT-generated text by default in the EU to comply with the EU AI Act, while keeping the feature optional in other regions. The proprietary textGrain method embeds detectable patterns in word choices, with detector access limited to selected researchers and organizations.
Mistral Large 4 reportedly tops a benchmark related to EU AI regulation compliance, highlighting how regional policy considerations are becoming a competitive factor for European AI models like Mistral's.
OpenAI 将在欧盟为 ChatGPT 和 Codex 生成的文本添加不可见水印以符合欧盟 AI 法案的透明度要求，同时发布名为 textGrain 的技术报告，但承认水印可被同义词替换等编辑削弱检测效果。
OpenAI is introducing invisible text watermarks called textGrain to ChatGPT and Codex for EU users to comply with the EU AI Act, with opt-in watermarking for API customers worldwide and detector access for approved researchers.
OpenAI announces its compliance approach to the EU AI Act's text provenance requirements, rolling out a phased text watermarking system called textGrain across the API, ChatGPT, and Codex while being transparent about detection limitations, and plans to open source the technology.
Aleph Alpha released Kolibri, an open-weight sovereign German/English Mixture-of-Experts LLM with 78B total parameters (~3.5B active per token), trained from scratch on German and Finnish infrastructure under Apache 2.0, with weights on Hugging Face and a 262k-token native context.
European banks, fintechs, and insurers are deploying AI agents that execute tasks like triggering payments while EU AI Act transparency rules (Article 50) apply but high-risk obligations remain delayed to 2027, leaving a regulatory gap alongside GDPR Article 22 and DORA. The author poses open questions on accountability, data residency, and whether Europe's documentation-heavy approach will slow deployment or become a trust advantage.
该论文对12款主流的EU AI Act合规检查工具（AIACCs）进行了首次实证评估，发现这些工具质量参差不齐，只能作为早期参考，过度简化关键义务且无法提供确定、可执行的指导，可能造成虚假的合规感。
This article analyzes how LLM watermarking, specifically SynthID-Text, impacts AI agent behavior by causing sampling drift that affects model refusals and tool calling, with implications for AI safety and regulatory compliance.
This paper presents a systematic analysis of the EU AI Act's high-risk requirements, deriving a list of AI-specific risk sources to bridge legal obligations with AI risk management practices.
A publicly available AI monitor tool that aggregates data on AI models, developers, countries, usage, performance, incidents, and the EU AI Act into an interactive environment for exploration and comparison.
Alibaba's Qwen team released Qwen3.8-Flash-Next, an open-weight preview of the Qwen4 architecture that activates only 6B parameters out of 125B to reduce inference costs and address hardware limitations.
The post discusses how the EU AI Act and the US's pro-acceleration stance create an accidental global experiment to see if Europe protected itself or regulated itself out of the AI frontier.
The article discusses how the stigma around AI use unfairly penalizes legitimate applications, referencing watermarking initiatives and transparency regulations from OpenAI, Anthropic, and the EU AI Act.
Coders have quickly developed workarounds to remove invisible watermarks from Claude-generated text, highlighting challenges in AI content detection and compliance with the EU AI Act.
Anthropic explains how Claude's invisible text watermarks, based on Google DeepMind's SynthID-Text, will work to comply with EU AI Act transparency requirements.
The EU AI Act mandates watermarking for AI-generated text to ensure detectability within the EU, even for outputs generated outside, with OpenAI planning to comply in future models like Astra.