Skip to content

Translation in progress Some entries still include text in the original language.

All updates

Indexed today 56 items

Page 26 Newer contentBack to latest

Oct 3Sat
  1. GitHub 安全公告 · Indexed · Original News51

    Five LLM-callable tools in Vibe-Trading disclosed with command execution, code injection, and SSRF vulnerabilities(original, opens in a new tab)

    A GitHub security advisory discloses that five LLM-callable tools in Vibe-Trading have command execution, code injection, and SSRF vulnerabilities, with BashTool and BackgroundRunTool both scoring 9.0 on CVSS v3.1. These tools are unconditionally registered into the auto-discovered tool registry under default configuration, the LLM can call them freely, and the container does not set a USER directive, so successful execution runs as uid=0 (root). BashTool passes LLM-generated commands verbatim to subprocess.run(shell=True) with no allowlist, escaping, or length limits; BackgroundRunTool executes the same kind of commands asynchronously, making execution harder to discover in access logs. The advisory also notes that even if the unauthenticated interface is fixed, attackers can still perform prompt injection on the Agent through malicious documents, turning legitimate callers into RCE channels.

    Why it was selected公告逐条给出 Vibe-Trading 五个 LLM 可调用工具的注入点、CVSS 评分与复现步骤,部署同类 Agent 的团队可据此对照检查工具注册与沙箱配置。

    1 reports ·1 sourcesView the event timeline and all reports
  2. GitHub 安全公告 · Indexed · Original News48

    SiYuan Note's MCP asset.upload has a workspace boundary bypass vulnerability, fixed in 3.8.1(original, opens in a new tab)

    SiYuan Note's MCP tool asset.upload lacks workspace boundary validation and can read files at arbitrary absolute paths; affected versions are 3.8.0 and below, fixed in 3.8.1. This tool, at kernel/mcp/tools/asset.go:195, only normalizes paths with filepath.Abs and does not perform IsSubPath or IsSensitivePath checks; subsequently, InsertLocalAssets will os.Open these paths and copy their contents into the workspace assets/ directory. An attacker can use prompt injection to make the Agent call this tool with files such as /Users/victim/.ssh/id_rsa, ~/.aws/credentials, or /etc/passwd as the files parameter, while the confirmation dialog only shows a category-level prompt and does not display the real source path, making it difficult for users to make an informed decision.

    Why it was selected披露思源笔记 MCP 工具缺失工作区边界校验,可被提示注入诱导读取任意本地文件,附补丁建议。

    1 reports ·1 sourcesView the event timeline and all reports
  3. Gizmodo · Indexed · Original News43

    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。

  4. Apollo Research · Indexed · Original News55

    Apollo Research CEO Marius Hobbhahn testifies before the U.S. Senate on misaligned AI(original, opens in a new tab)

    Apollo Research founder and CEO Marius Hobbhahn testified before a subcommittee of the U.S. Senate Committee on Homeland Security and Governmental Affairs, discussing risks from AI scheming (models deliberately deceiving humans, hiding true capabilities and intentions to pursue their own goals). He stated that capabilities are advancing faster than alignment: Claude accelerated small model training code by 3x in May 2025 and reached 52x in April 2026; Claude currently handles 26% of Anthropic's internal AI research and development, up from under 1% in February; an unreleased OpenAI model ran about 10,000 agents for 88 hours, and machine verification solved the Navier–Stokes problem.

    Why it was selectedApollo Research's CEO, testifying before Congress, proposed four regulatory recommendations and provided specific data on evaluation awareness and declining chain-of-thought readability.

    5 reports ·4 sourcesView the event timeline and all reports
  5. Neel Nanda · Indexed · Original X22

    AI 安全倡导者被指"心理战"实为资金隐秘的舆论操作(original, opens in a new tab)

    Neel Nanda 指出,指控 AI 安全倡导者是"资金隐秘的心理战"的一方,本身正是资金隐秘、意图误导公众的舆论操作。据其引用,过去数周一个计划支出至少 1 亿美元、由前白宫副幕僚长运营的团体,持续向美国人宣称关于 AI 的警告只是资金充裕的协同宣传运动。他认为围绕安全的公共讨论固然重要,但这类操作无助于对话。

    1 reports ·1 sourcesView the event timeline and all reports
  6. FAR.AI · Indexed · Original X8

    Far AI 招聘 AI 安全研究与红队岗位(original, opens in a new tab)

    Far AI 正在招聘,称团队快速扩张,需要投身 AI 安全这一重要问题的人才。岗位既包括研究员、红队测试人员和工程师等技术角色,也涵盖运营、项目管理和活动筹办等非技术角色。申请链接见评论区,并鼓励转发给合适人选。

    1 reports ·1 sourcesView the event timeline and all reports
  7. The Midas Project · Indexed · Original X11

    美中AI安全合作能否实现(original, opens in a new tab)

    “美国赢不了中国,中国也赢不了美国。我们都打开了潘多拉魔盒……认为中国不愿为人类利益参与[AI安全保障合作]——我不同意。这是一个有待验证的命题。”

    1 reports ·1 sourcesView the event timeline and all reports
  8. Zenity · Indexed · Original X13

    OWASP、MITRE ATLAS 与 NIST 各自构建智能体 AI 安全框架(original, opens in a new tab)

    OWASP、MITRE ATLAS 和 NIST 在过去一年各自建立了面向智能体 AI 的专门框架,独立得出同一结论:智能体需要独立的安全类别。Zenity 发布《企业智能体 AI 安全买家指南》,拆解了这一趋势的成因,并列出评估平台时真正重要的能力。

    1 reports ·1 sourcesView the event timeline and all reports
  9. Zenity · Indexed · Original News37

    Zenity Labs launches AI Total: sandbox detonates Agent Skills, uncovers credential-stealing campaign with approximately 1.7 million installs(original, opens in a new tab)

    Zenity Labs released a free service called AI Total that uses a Detonation Chamber to run Agent Skills in a sandbox, recording their behavior and comparing it against claims. Before release, it detonated thousands of public Skills and found dozens of malicious samples, one of which—a credential-stealing campaign distributed via Vercel skills.sh—reached approximately 1.7 million installs; over 30% of malicious Skills instructed agents like Claude Code and OpenClaw to download and run attacker files.

    2 reports ·2 sourcesView the event timeline and all reports
  10. Dark Reading · Indexed · Original News13

    RemoteThreat launches attack and defense drill platform, betting that security teams need to test scenarios after defenses fail(original, opens in a new tab)

    RemoteThreat comes out of stealth mode with $7 million in pre-seed funding, launching an integrated offensive cybersecurity operations platform that uses AI to help enterprise and government teams test how they perform after their defenses are breached. The platform builds tools from scratch each time, can bypass Tier 1 EDR vendors and security controls, supports fully manual, AI-assisted, or hybrid modes, and allows organizations to use their own AI models. CEO Chris Thompson predicts that as frontier models advance rapidly, the traditional penetration testing model will cease to exist within two years.

    1 reports ·1 sourcesView the event timeline and all reports
  11. threatfrontier.com · Indexed · Original News55

    Anthropic 修复 Claude Desktop macOS 漏洞:Cowork 文件夹内文件可执行命令(original, opens in a new tab)

    Anthropic 发布高危公告 GHSA-v234-4jrq-mgg6,修复 Claude Desktop on macOS 的一个漏洞:被攻陷或被提示注入的 Agent 写入 Cowork 共享文件夹的某类文件,在用户从 Claude Desktop 打开时可在 Mac 上执行命令。受影响版本为 1.1.3918 至 1.15962.0 之前,修复版本为 1.15962.0,公告于 2026 年 9 月 25 日发布且未列出 CVE 编号。问题根源是 Claude Desktop 的禁止执行文件类型清单遗漏了 macOS 打开即执行的一类文件,被归类为 CWE-184 不完整禁止输入清单,CVSS 4.0 评分 8.5,属本地攻击向量、低复杂度、无需权限、需被动用户交互。

    Why it was selected材料给出受影响版本区间、修复版本与 CVSS 评分,并说明攻击需要被提示注入的 Agent 先写入文件,便于运维核对升级状态。

  12. Financial Times · 人工智能 · Indexed · Original News50

    Tencent signs agreement with Oracle to lease 100,000 chips to accelerate AI layout(original, opens in a new tab)

    Tencent signed its largest overseas leasing agreement with US cloud service provider Oracle, reaching a five-year leasing arrangement this year covering multiple Oracle data centers in Southeast Asia, thereby obtaining about 100,000 advanced AI chips that cannot be obtained in China, to advance its AI model and agent tool development. Two people familiar with the matter said the agreement is valued at about 7 billion US dollars, with a down payment of about 30%, and caused Tencent's second-quarter free cash flow to be negative 13.8 billion yuan (about 2 billion US dollars), the first quarterly negative value in more than a decade. Tightening US export controls and domestic supply shortages have prompted Chinese tech companies to turn to overseas computing power; such leasing agreements are permitted under US rules and may also allow companies to use cutting-edge Nvidia processors that cannot be obtained in China. Tencent's second-quarter capital expenditure increased 176% year-on-year to 53 billion yuan (about 7.9 billion US dollars), for AI infrastructure and prepayments for computing resources; the company also launched the embedded agent Xiaowei in WeChat and built an enterprise-level AI agent matrix led by WorkBuddy.

    1 reports ·1 sourcesView the event timeline and all reports
  13. Financial Times · 人工智能 · Indexed · Original News11

    AI 真能终结人类吗?牛津大学 Toby Ord 谈 AI 生存风险焦虑(original, opens in a new tab)

    牛津大学 AI Governance Institute 高级研究员、《The Precipice: Existential Risk and the Future of Humanity》作者 Toby Ord 在 FT 播客中讨论 AI 生存风险。高关注度 AI 安全事件与业内人士对 AI 快速进展的警告,引发了新一轮对人类未来的担忧;节目追问 AI 系统能力不断增强之际,能否有效监管这项技术,还是已经错过把 AI 精灵收回瓶中的时机。

    1 reports ·1 sourcesView the event timeline and all reports
  14. The Guardian · 人工智能 · Indexed · Original News31

    澳大利亚通信部长 Anika Wells 因未成年人社交媒体禁令入选 Time100 新星榜(original, opens in a new tab)

    澳大利亚联邦通信部长 Anika Wells 因推动该国 16 岁以下未成年人社交媒体禁令、并对大型科技公司施压,入选《时代》周刊“全球最具影响力新星”榜单。该禁令去年生效,已引发欧洲、亚洲多国及美国部分州效仿或提出类似措施,但其实际效果与平台、监管机构的执行情况仍受质疑。Wells 目前正推进数字注意义务框架,允许用户选择退出推送危险与分裂内容的算法。

    1 reports ·1 sourcesView the event timeline and all reports
  15. Transformer · Indexed · Original News36

    Human oversight cannot defuse the risk of AI in war(original, opens in a new tab)

    Transformer published an article arguing that keeping humans in the decision loop does not eliminate the risks of AI involvement in military decisions, and that the real danger lies in human operators who over-rely on AI. The article cites a CNN report from September 18 that, during a war against Iran, the U.S. military prepared to board a Chinese cargo ship in the Middle East suspected of carrying nuclear components, based on a completely erroneous intelligence assessment produced by an analyst with the help of AI; aircraft had already taken off, and some said the incident nearly triggered war between two nuclear powers. Jacquelyn Schneider of Stanford University's Hoover Institution said AI makes analysts more confident in their assessments while seeing less of the data sources. The article also mentions Israel's Lavender system, which had an error rate of up to 10%, with personnel often rubber-stamping its decisions; Palantir's Maven Smart System was accused of being involved in the first-day bombing of an elementary school in Minab, killing more than 150 people, possibly due to outdated data, after which Palantir upgraded the system to require re-review of underlying intelligence.

    2 reports ·2 sourcesView the event timeline and all reports
  16. The Guardian · 人工智能 · Indexed · Original News53

    Guardian test: some chatbots remove religious clothing from images, Claude refuses(original, opens in a new tab)

    The Guardian tested ChatGPT, Grok, Gemini, and Claude, asking them to remove the hijab from a woman in an AI-generated image. ChatGPT and Grok did it directly; Claude said it has no image editing or generation capabilities and would not modify photos to remove a hijab; Gemini initially refused on the grounds that it should not digitally alter someone's appearance without consent, but after being asked to make the woman look more "Western," it generated an image without the hijab. The test was prompted by French National Rally MP Julien Odoul posting a modified photo of a real Muslim woman on X with the caption "LIBERTÉ FRANÇAISE." Grok refused to remove a dress from an AI-generated woman but called removing a hijab a "relatively straightforward clothing or hairstyle modification" and said it was willing to make a "less explicit" version; ChatGPT also refused to remove the dress but acknowledged that exposing hair for a hijab-wearing person could constitute an infringement of privacy or religious practice. OpenAI, Google, xAI, and Anthropic all declined to comment.

    1 reports ·1 sourcesView the event timeline and all reports
  17. AWS Security · Indexed · Original News57

    AWS 披露 Loom for AWS 三个漏洞,1.7.0 版本完成修复(original, opens in a new tab)

    AWS 披露并修复了开源 AI 智能体编排平台 Loom for AWS 中的三个漏洞,建议升级到 1.7.0 并对派生代码打补丁。CVE-2026-103956 是认证绕过(CWE-306、CWE-1188),在未配置身份提供方的部署中,任意网络客户端可经应用 API 获得智能体控制面的完全管理权限,包括注册工具服务器、读取已存集成凭据和改写托管智能体角色的 IAM 策略,该问题已在 2026-08-04 发布的 1.6.1 中修复。CVE-2026-103957 涉及 OAuth2 发现处理(CWE-918、CWE-201),拥有 mcp:write 或 a2a:write 权限的已认证用户可配置发现 URL,使后端把 OAuth2 客户端密钥或其他用户访问 token 发送到第三方端点,1.6.1 仅阻断内网地址访问,1.7.0 才完全修复。 第三项 CVE-2026-103958 同样影响 1.7.0 之前的版本,涉及 MCP/A2A 连接中的服务端请求伪造,可能触及内部服务与临时凭证;公告建议统一升级 1.7.0,未称已被在野利用。

    Why it was selectedAWS 披露 Loom 三个 CVE 的成因与修复版本,并给出升级前后的凭据轮换与权限收紧步骤,可对照自查部署。

  18. Redwood Research · Indexed · Original News33

    Capability research also expands the safety-usefulness Pareto frontier(original, opens in a new tab)

    Redwood Research proposes that defining safety research as "improving deployment safety without significantly sacrificing usefulness" would count almost all capability research as safety research, for example reasoning performance optimization making weaker, safer models more widely used. The authors argue this criterion is insufficient: developers must choose a point on the Pareto frontier, safety research typically guides them to choose higher safety, while capability research does the opposite, because dangerous AI is more useful. The authors also note that in a future scenario where political will is much higher than today, some capability research could become an effective way to improve safety.

    1 reports ·1 sourcesView the event timeline and all reports
  19. GitHub 安全公告 · Indexed · Original News51

    DeepTutor versions before 1.4.10 have an MCP tool authorization bypass vulnerability(original, opens in a new tab)

    DeepTutor versions before 1.4.10 have an authorization bypass vulnerability (CVE-2026-58168), where low-privilege users can invoke unrestricted MCP tools. The cause is that the allowed_mcp_tools function in deeptutor/multi_user/tool_access.py returns None when mcp_tools is omitted from user authorization, rather than a denial result. Attackers or prompt-injected content in user sessions can enumerate and invoke any configured MCP tool, including filesystem, shell, and browser servers, thereby gaining unauthorized access to sensitive resources in the deployment. The advisory was added to the GitHub Advisory Database on June 30, 2026, and updated on October 2.

    Why it was selected公告给出了漏洞函数与受影响版本,部署 DeepTutor 多用户环境的团队可据此判断自身 MCP 工具权限是否被绕过。

    1 reports ·1 sourcesView the event timeline and all reports
  20. OECD.AI(政策与事件监测) · Indexed · Original Date unknownNews29

    弥合 AI 评估差距的五步路线图(original, opens in a new tab)

    OECD.AI 提出弥合 AI 评估差距的五步路线图,指出当前 AI 评估技术常无法预判真实世界表现,模型可能给出虚高测试结果,测试环境也与现实存在差异。2026 年国际 AI 安全报告由 100 多名专家编写、30 多个国家和多边机构支持,提出"AI 证据困境"。路线图第一步是平衡标准化与定制化,在 2026 年印度 AI 影响峰会上多家公司承诺改进多语言与情境化评估。

  21. OECD.AI(政策与事件监测) · Indexed · Original Date unknownNews17

    HAIP 如何将透明度从合规负担转变为竞争优势(original, opens in a new tab)

    Salesforce 参与 OECD 广岛 AI 进程(HAIP)报告框架,认为该框架为智能体 AI 时代提供了跨司法管辖区的通用合规语言。文章指出,当前各国风险与透明度定义分歧造成监管碎片化,中小企业在多市场合规中处于劣势,而统一的分级要求框架可降低门槛。Salesforce 的 Agentforce 平台已在欧盟、美国、日本等地部署,需为每个司法管辖区单独准备合规文档。

  22. Snyk Labs(AI Threat Labs,含原 Invariant Labs) · Indexed · Original Date unknownNews36

    Snyk Labs 解读 MAESTRO:面向 Agentic AI 生态的分层威胁建模框架(original, opens in a new tab)

    Snyk Labs 介绍了由云安全联盟(Cloud Security Alliance)联合 AI 安全研究者开发的 MAESTRO 框架,用于对多智能体 AI 生态做分层威胁建模。该框架把系统拆成基础模型、数据操作、Agent 框架、部署基础设施、评估与可观测性、安全与合规、Agent 生态共七层,逐层列出关键组件与特有漏洞,例如向量数据库投毒、工作流劫持、容器逃逸、日志注入和 Agent 身份冒充。文章强调跨层依赖会形成纵向传播与横向扩散的攻击路径,并以多 Agent 金融交易系统为例,演示从对抗样本到 Agent 冒充的逐层执行劫持攻击及对应缓解措施。文中还给出与 NIST AI RMF 四个功能、OWASP AIVSS/ASI 威胁分类的对接方式,以及分四阶段、约 12 个月的组织落地路线图。

  23. Snyk Labs(AI Threat Labs,含原 Invariant Labs) · Indexed · Original Date unknownNews51

    Snyk 推出 Agent Scan - Skill Inspector,免费扫描 Agent 技能中的恶意行为(original, opens in a new tab)

    Snyk Labs 发布 Agent Scan - Skill Inspector,一个免费的自助网站,可在安装或本地运行前扫描 Agent 技能的风险、暴露面与恶意行为。团队对主流市场的 3,984 个技能做了分析,确认 76 个恶意技能,13.4% 的技能至少含一个严重级别问题,36.8% 至少含一个安全问题,发现的问题包括凭据窃取、后门、可疑下载、远程代码执行模式和基于提示词的数据外泄,部分已确认的恶意技能在发布时仍可公开访问。Agent Scan 将技能视为代码与自然语言指令(如 SKILL.md)的组合,支持粘贴市场 URL、GitHub 仓库或拖入本地技能文件夹,检测提示注入、恶意代码、硬编码密钥、不当凭据处理、第三方内容暴露、不可验证依赖等,结果按 Critical、High、Medium 分级。

    Why it was selectedSnyk 对 3,984 个 Agent 技能的大规模扫描数据,为评估技能市场供应链风险提供了可参照的基线。

  24. Snyk Labs(AI Threat Labs,含原 Invariant Labs) · Indexed · Original Date unknownNews43

    Snyk Labs 披露 OpenClaw 沙箱两处绕过:策略未合并与 TOCTOU 竞态(original, opens in a new tab)

    Snyk Labs 对 OpenClaw 做安全评估,发现两处沙箱绕过,结果都是配置里看似存在的沙箱边界在运行时并未生效。第一处是 /tools/invoke HTTP 端点在构建和过滤工具列表时没有合并 sandbox 策略,只叠加了 Profile、Global、Agent、Group、Subagent 五层策略,导致 browser、gateway、nodes 等沙箱内本应禁用的工具仍可调用,任何持有有效 gateway 凭据的远程集成或插件都能借此提权;作者已向 OpenClaw 仓库提交修复,使该处理器按 sessionKey 解析沙箱上下文并合并允许与拒绝列表。作者认为在 Node.js 中无法用 openat(2) 配合目录文件描述符彻底修复,最终把文件操作移入沙箱的 Docker 容器内执行。

  25. 日本 AI 安全研究所(J-AISI) · Indexed · Original Date unknownNews48

    日本 AI 安全研究所用 ExploitBench 评测 Claude Opus 4.8 的网络安全能力(original, opens in a new tab)

    日本 AI 安全研究所(J-AISI)用基于已知漏洞的 ExploitBench 评测了 Anthropic 的 Claude Opus 4.8 的网络安全能力,结论是其整体能力与前代 Opus 4.7 相当或更强。评测覆盖 41 个 V8 已知漏洞,在 V8 Security Sandbox 环境中每个任务只跑一次(seed 1,最多 300 轮、最多 5 小时),并在关闭实时网络安全护栏的条件下进行,因此不代表常规部署环境下的表现。在可对比的 40 个任务中,Opus 4.8 有 15 个达到更高 Tier、23 个持平、2 个更低,最多任务停留在对应沙箱内数据读写的 T3。其中一个任务(crbug-1509576)出现了 T1 所需的 pc_control,但未获得另一项 ace,且模型未按基准预设的逐级路径推进,追试也未能复现该结果;作者判断这很可能只是本次评估条件下才出现的结果。

    Why it was selected日本 AI 安全研究所用 ExploitBench 评测 Claude Opus 4.8 的漏洞利用能力,并说明为何单看 Score 与 Tier 会误判模型能力。

  26. 日本 AI 安全研究所(J-AISI) · Indexed · Original Date unknownNews54

    日本 AI 安全研究所评测:GLM-5.2 的漏洞利用能力仍落后于 Opus 系模型(original, opens in a new tab)

    日本 AI 安全研究所(J-AISI)用 ExploitBench 评测开源权重模型 GLM-5.2 的漏洞利用开发能力,发现其部分任务展现高级能力,但整体仍落后于 Opus 4.7 和 Opus 4.8。评测覆盖 V8 引擎的 41 个已知漏洞任务,GLM-5.2 平均得分 2.80,Opus 4.7 为 3.63,Opus 4.8 为 5.22;GLM-5.2 最高只到 T3 档,而 Opus 4.8 在部分任务上达到 T2 和 T1。在三个模型都达到 T5 的 36 个任务上,到首次达到 T5 为止的 token 费用,GLM-5.2 为 3.71 美元,低于 Opus 4.7 的 4.60 美元和 Opus 4.8 的 7.34 美元。在 10 个高风险任务中,加入 nudge 后 GLM-5.2 停在 T5 的任务从 5 个变为 6 个,平均分仍为 3.0,未实现更高级的漏洞利用。

    Why it was selected日本 AI 安全研究所用 ExploitBench 对比 GLM-5.2 与 Opus 系模型的漏洞利用能力,给出了分档得分与成本数据。

  27. Lakera Blog · Indexed · Original Date unknownNews8

    AI 安全不再是一个单一问题:Lakera 提出 AI Defense Plane 统一防护员工、应用与智能体(original, opens in a new tab)

    Lakera 提出 AI Defense Plane,主张把员工、应用与智能体三层 AI 风险纳入同一控制平面,而非各自部署点状防护。该框架强调对 AI 使用端到端可见、在运行时执行策略,并跨层关联信号,同时持续在真实与对抗条件下测试系统行为。其背景是 AI 已从生成输出转向执行动作,智能体常以委派权限调用工具,提示注入与意外泄露出现在传统控制难以检查的流程中。

  28. Lakera Blog · Indexed · Original Date unknownNews44

    Lakera 扫描 npm 发现 30 个包随 Claude Code 权限文件泄露凭据(original, opens in a new tab)

    Lakera 构建 TypeScript 服务监控 npm 注册表的 CouchDB 变更流,抓取新发布包的 tarball 检查是否含 .claude/settings.local.json。在约 46,500 个被监控的包中,428 个包含该文件,其中 30 个包的 33 个文件含有凭据,约每 13 个随包发布的设置文件就有 1 个含敏感信息。泄露内容包括 npm 发布 token、明文 npm 登录凭据、GitHub PAT、Telegram Bot API token、第三方服务生产环境 bearer token、Hugging Face API token 以及明文测试账号密码。

  29. Lakera Blog · Indexed · Original Date unknownNews16

    Check Point 与 Google Cloud 合作,为 Gemini Enterprise Agent Platform 上的 AI 智能体提供结果控制(original, opens in a new tab)

    Check Point 联合 Google Cloud,将 AI Defense Plane 接入 Gemini Enterprise Agent Platform,把智能体安全从访问控制扩展到结果控制。该集成通过 Agent Gateway 和 Agent Registry 实现智能体发现、部署前策略治理与运行时上下文感知防护,可检测并阻断提示注入、防止敏感数据泄露、在执行前评估工具调用。集成将于 2026 年 6 月下旬开放。

  30. Lakera Blog · Indexed · Original Date unknownNews17

    AI 已不再请求许可:企业安全团队是否察觉自主智能体带来的风险(original, opens in a new tab)

    AI 正从"建议型"转向"行动型",可自主检索内部数据、调用 API、修改记录并触发工作流,而传统安全工具无法在推理层检查其意图。Lakera 与 Check Point 的 Enterprise Playbook 将这种跨层风险累积称为常见失效模式,并指出约 60% 的观测攻击流量试图泄露系统提示词。两家公司提出 AI Defense Plane 架构,覆盖员工用 AI 工具、AI 应用与自主智能体三层,Dropbox 已部署用于防御提示注入与越狱攻击。

  31. Lakera Blog · Indexed · Original Date unknownNews4

    Lakera 与 Check Point 推出 READY OR NOT:AI 企业安全 5 场系列网络研讨会(original, opens in a new tab)

    Lakera 与 Check Point 联合推出 READY OR NOT 五场系列网络研讨会,聚焦企业 AI 安全就绪的实践落地。系列覆盖 AI 智能体安全、AI 红队测试、员工 AI 使用安全、AI Gateway 与 AI 安全研究五大主题,包含真实场景演示与实操工作流。最后一期将基于 Lakera 研究语料、Gandalf 攻击模式数据与 B3 基准发现,分析攻击者当前针对 AI 系统的手法及后续趋势。

  32. Pillar Security · Indexed · Original Date unknownNews63

    Pillar Security discloses a Dolt MCP authentication bypass vulnerability; unauthenticated callers can execute database tools(original, opens in a new tab)

    Pillar Security discovers an authentication bypass vulnerability in the DoltHub remote Dolt MCP server, affecting versions 0.3.1 through 0.3.6, fixed in 0.3.7. The vulnerability is caused by the combination of two defects: the JWT middleware writes a 401 response when the token is invalid but lacks a return, so processing continues downward; meanwhile, the mark3labs/mcp-go default session manager only validates the format of the Mcp-Session-Id, not whether it was issued by the server. An attacker can therefore reach the tool distribution stage with a forged JWT and a self-made session ID, execute operations using the server's own database credentials, and the results are appended to the 401 response body, so logs that only monitor status codes will mistakenly judge the attack as failed. The impact is limited to remote HTTP transports with JWT authentication enabled; local stdio is not affected. The operations that can be executed depend on the database permissions of the MCP server account, including reading, modifying, or deleting data.

    Why it was selectedDiscloses the full chain of the Dolt MCP authentication bypass vulnerability and the fixed versions; teams deploying MCP services can use this to check whether tools are still executed after authentication failure.

  33. Pillar Security · Indexed · Original Date unknownNews46

    Pillar 披露 Gemini CLI 攻击链:GitHub issue 提示注入可致 GCP 项目权限失陷(original, opens in a new tab)

    Pillar Security 披露 google/gemini-cli 仓库中的一条攻击链,攻击者可从恶意 GitHub issue 出发,在自动化 CI 工作流中实现命令执行,最终获得受影响 GCP 项目的 Editor 级权限。攻击链的关键在于工作流使用的顶层 coreTools 配置在当前 CLI(0.46.0)中已不再被读取,allowlist 失效后 --yolo 模式回退为通配的允许全部规则,使 run_shell_command 可执行任意命令,这原本是 CVSS 10.0 的 GHSA-wpqr-6v78-jr5g 试图修复的问题。Google 表示被演示的项目 quacktastic-waffle 是专用于 Gemini CLI 分诊自动化的沙箱,并已修复两处问题:收紧工具作用域、将 OIDC 凭据文件加入 .geminiignore,并把该角色限制到单个服务账号资源。

  34. Pillar Security · Indexed · Original Date unknownNews41

    Pillar Security 披露 Grafana MCP 会话伪造与 SSRF 漏洞(original, opens in a new tab)

    Pillar Security 在 Grafana MCP 中发现两个可组合成严重级别攻击链的安全问题,受影响部署中未认证攻击者可调用 MCP 工具并借服务器网络位置访问内部服务。第一个问题允许可达调用者用自行生成的会话 ID 通过校验,无需任何凭据即可调用 tools/list 和 grafana_api_request,从而以服务器配置的 Grafana 服务账号身份发起请求;Grafana 在 v1.1.0 中加入可选的 bearer token 认证,未认证请求会在工具执行前返回 401。第二个问题中 grafana_api_request 的 X-Grafana-URL 允许调用者指定出站请求目的地,虽然跨主机时不再附带服务账号 token,服务器仍会发出请求并回传响应,构成 SSRF,编号 CVE-2026-19516,CVSS 9.1。

  35. Pillar Security · Indexed · Original Date unknownNews33

    EU AI Act 高风险义务延期至 2027:审计窗口期该做什么(original, opens in a new tab)

    EU AI Act 的高风险义务经 Digital Omnibus 延期至 2027 年 12 月 2 日,但第 50 条透明度义务已于 2026 年 8 月 2 日生效,违规最高可罚 1500 万欧元或全球年营收 3%。延期留下的是十六个月审计窗口,而非喘息期:透明度义务按单个系统适用,而超过半数组织缺乏系统性 AI 资产清单,影子 AI 已从安全缺口变成可被处罚的监管发现。

  36. Pillar Security · Indexed · Original Date unknownNews25

    2026 年企业级 AI 智能体与 GenAI 应用最佳 AI 红队供应商对比(original, opens in a new tab)

    2026 年企业级 AI 红队供应商的评判标准,是测试已部署系统而非仅测试底层模型、每次发布都重跑测试、并给出执行证据而非分数。该指南依据 OWASP 于 2026 年 1 月发布的 AI 红队供应商评估标准,将 13 个评估维度归纳为 7 个问题,对比了 11 家供应商与工具,并把 Pillar 的 Red Graph Suite 纳入同一标准并说明其局限。文中指出,平均一次成功攻击需 5 轮对话,生产环境中 90% 的成功攻击导致数据泄露。

  37. Pillar Security · Indexed · Original Date unknownNews12

    Pillar Security 获 Latio 2026 AI 安全市场报告评为 AI 安全技术创新者(original, opens in a new tab)

    Latio 在 2026 AI 安全市场报告中把 Pillar Security 评为 AI 安全技术创新者,并将其列入覆盖端点、SaaS 与第一方智能体的 Broader AI Security 类别。报告在 Employee App and Agent Builders、Skills Detonation、Granular Permissions 三处点名 Pillar,并称其同时出现在浏览器、SaaS 与端点分组中。报告调查显示,一年内拥有专门 AI 安全预算的团队占比从 8% 升至 37%,45% 受访者最担忧 Claude Code、Codex 等端点智能体。