←搜索结果

按匹配度 · 共 85 条结果 清除搜索

星期二

  1. TLDR AI

    麦肯锡:仅 1% 的企业达到 AI 成熟阶段

    麦肯锡研究显示,尽管 92% 的企业计划未来三年增加 AI 投资,但仅 1% 的企业领导者认为其组织的 AI 部署已达到成熟阶段。员工广泛使用 AI 的可能性是领导者认知的三倍,已有 13% 的员工将 AI 用于超过 30% 的日常任务。

    resents a significant evolution from simple chatbots to truly autonomous AI workers that can complete entire business processes without human intervention. Source: OpenAI Summary: AWS announced Amazon Bedrock AgentCore, a comprehensive platform for deploying and operating AI agents at enterprise scale. The service includes seven core capabilities for secure agent deployment, new marketplace listings for AI agent tools, and a $100 million investment in the AWS Generative AI Innovation Center to a

    匹配字段正文

星期三

  1. Johann Rehberger / Embrace The Red精选

    GitHub Copilot 提示词注入致远程代码执行(CVE-2025-53773)

    安全研究者 Johann Rehberger 披露 GitHub Copilot 与 VS Code 中的提示词注入漏洞 CVE-2025-53773,攻击者可通过源码文件、网页、GitHub issue 或 MCP 返回内容中的注入指令,向 ~/.vscode/settings.json 写入 "chat.tools.autoApprove": true,使 Copilot 进入 YOLO 模式并关闭全部用户确认,进而执行终端命令实现远程代码执行。该漏洞影响 Windows、macOS 和 Linux,作者演示了下载恶意软件、接入远程命令控制服务器,以及通过感染文件在 Git 项目间传播的 AI 病毒思路。作者于 2025 年 6 月 29 日报告,微软在 8 月补丁星期二发布修复;文中还提到 .vscode/tasks.json 和伪造 MCP 服务器等其他可导致代码执行的攻击面。

    lso have discovered it in parallel. Thanks to the members of the MSRC and product team for the help in getting it mitigated. Conclusion This is another example of how an AI agent might not stay in its box! By modifying its own environment GitHub Copilot can escalate privileges and execute code to compromise the developer’s machine. It’s a not uncommon design flaw in agentic systems as I have discovered. Keep looking out for such design flaws, these should be easily caught during threat modeling.

    匹配字段正文

星期六

  1. Cato AI Labs精选

    Aim Labs 披露 M365 Copilot 零点击漏洞 EchoLeak 的完整攻击链

    Aim Labs 披露了 Microsoft 365 Copilot 中的零点击漏洞 EchoLeak,攻击者只需向受害者发送一封邮件,就能在用户无感知的情况下外泄 Copilot 上下文中的敏感数据。该攻击链由三处漏洞组成,依次绕过 XPIA 提示注入分类器、外部链接与图片的脱敏处理,并借助 SharePoint 和 Teams 的 URL 绕过 CSP 限制。Aim Labs 将核心手法称为 LLM Scope Violation,认为同类设计缺陷也可能存在于其他基于 RAG 的聊天机器人和 AI 智能体中,并称 Microsoft 已确认没有客户受到影响。

    nst such novel vulnerabilitiesAim Labs is not aware of any customers being impacted to date. TL;DR Aim Security discovered “EchoLeak”, a vulnerability chain that exploits design flaws typical of RAG Copilots, allowing attackers to automatically exfiltrate any data from M365 Copilot’s context, without relying on specific user behavior. The main chain is made up of three distinct vulnerabilities, but Aim Labs has identified additional vulnerabilities in its research process. Attack Flow What is a

    匹配字段正文

星期四

  1. Lilian Weng:Lil'Log精选

    Lilian Weng 长文梳理强化学习中的奖励劫持

    Lilian Weng 在 Lil'Log 发布长文,系统梳理强化学习中的奖励劫持问题,即智能体利用奖励函数的缺陷或歧义获取高分却未真正完成任务。文章按 RL 任务、LLM 任务和现实生活分类列举实例,并讨论 RLHF 中代理奖励与真实奖励的差距、LLM 作为评估者的位置偏差,以及对抗奖励函数、解耦审批、异常检测等缓解思路。作者指出,随着模型能力增强,奖励劫持更易出现,而针对 RLHF 和 LLM 的实用缓解研究仍然有限。

    cking examples in real life; Why does Reward Hacking Exist?; Hacking RL Environment; Hacking RLHF of LLMs Reward hacking occurs when a reinforcement learning (RL) agent exploits flaws or ambiguities in the reward function to achieve high rewards, without genuinely learning or completing the intended task. Reward hacking exists because RL environments are often imperfect, and it is fundamentally challenging to accurately specify a reward function. With the rise of language models generalizing to

    匹配字段正文

星期二

  1. Zyphra Research

    Zyphra 发布 5 万亿 token 数据集 Zyda-2,用 NVIDIA NeMo Curator 将处理时间从 3 周缩短到 2 天

    Zyphra 发布 Zyda-2,一个约 5 万亿 token 的宽松许可数据集,由 DCLM、FineWeb-Edu、Zyda-1 和 Dolma v1.7 的 Common Crawl 部分经跨数据集去重与模型过滤组合而成,已在 HuggingFace 上开放。借助 NVIDIA NeMo Curator 的 GPU 加速,数据处理从 CPU 上的 3 周缩短到 2 天,总拥有成本降低 2 倍。Zyphra 称用 Zyda-2 退火训练的模型在 MMLU、Hellaswag 等聚合评测上优于 DCLM 和 FineWeb-Edu,其 Zamba2 系列模型即基于该数据集的早期版本训练。

    the first album was released, Shaw and Deuce wasted no time with new material, releasing the "Beware Soul Snatchers" single within weeks of its appearance. An even denser slab of inboard studio trickery, "Soul Snatchers" was the clearest example to date of the group's evolving sound, though further recordings failed to materialize. ‍ 1 Leaving Perdition 8:10 2 Intermodal 3:39 3 Lunar White 3:25 4 She's A Black Belt 3:14 5 Rohm 1:09 6 Radiolucence (Version) 5:31 7 Deepest Laws 14:14 No comments:

    匹配字段正文