惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Google DeepMind News
Google DeepMind News
F
Fortinet All Blogs
阮一峰的网络日志
阮一峰的网络日志
Apple Machine Learning Research
Apple Machine Learning Research
爱范儿
爱范儿
WordPress大学
WordPress大学
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
J
Java Code Geeks
罗磊的独立博客
S
SegmentFault 最新的问题
V
V2EX
V
Visual Studio Blog
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
美团技术团队
博客园 - 三生石上(FineUI控件)
Stack Overflow Blog
Stack Overflow Blog
Y
Y Combinator Blog
MyScale Blog
MyScale Blog
D
Docker
Google DeepMind News
Google DeepMind News
Blog — PlanetScale
Blog — PlanetScale
M
Microsoft Research Blog - Microsoft Research
Martin Fowler
Martin Fowler
S
Secure Thoughts
B
Blog
cs.CL updates on arXiv.org
cs.CL updates on arXiv.org
www.infosecurity-magazine.com
www.infosecurity-magazine.com
Recent Announcements
Recent Announcements
MongoDB | Blog
MongoDB | Blog
C
Cisco Blogs
C
CERT Recently Published Vulnerability Notes
T
True Tiger Recordings
GbyAI
GbyAI
P
Proofpoint News Feed
P
Privacy International News Feed
Jina AI
Jina AI
The Cloudflare Blog
I
Intezer
AWS News Blog
AWS News Blog
Hacker News - Newest:
Hacker News - Newest: "LLM"
S
Security Archives - TechRepublic
NISL@THU
NISL@THU
The Register - Security
The Register - Security
Recent Commits to openclaw:main
Recent Commits to openclaw:main
P
Palo Alto Networks Blog
S
Schneier on Security
L
LINUX DO - 热门话题
C
CXSECURITY Database RSS Feed - CXSecurity.com
Security Latest
Security Latest
C
Cybersecurity and Infrastructure Security Agency CISA

The Register - Security

Trump Mobile site leaks customer data as phone finally ships Cisco used AI to write security incident reports, with mixed results Dems slam Trump cyber cuts amid ballroom, Jan. 6 'slush fund' Threat hunters find Google API keys still usable 23 minutes after deletion HackerOne takes an axe to its bug bounty rewards 46k plaintext passwords pwned in Myspace93 breach Cisco serves up yet another perfect 10 bug with Secure Workload admin flaw Zombie user account let hackers control the city’s water Even Claude agrees: hole in its sandbox was real and dangerous GitHub says internal repos exfiltrated after poisoned VS Code extension attack London's police asked Big Tech for comms data over 700,000 times last year Microsoft shuts down illegal code-signing operation used by ransomware crims to mask their malware America's top cyber-defense agency left a GitHub repo open with passwords, keys, tokens – and incredibly obvious filenames America's top cyber-defense agency left a GitHub repo open with with passwords, keys, tokens – and incredibly obvious filenames Clear your calendar, Drupal user: You have a critically urgent patch to install Clear your calendar, Drupal user: You have a critically urgent patch to install Do fear the Reaper - stealer swipes macOS users' passwords, wallets, then backdoors them Shai-Hulud copycat hits another npm package Linux kernel flaw opens root-only files to unprivileged users TanStack weighs invitation-only pull requests after supply chain attack NGINX Rift attackers waste no time targeting exposed servers Poland directs officials to ditch Signal in favor of 'secure' state-developed alternative F-35 software delays leave UK buying time with US glide bombs Mozilla warns UK: Breaking VPNs will not magically fix Britain's age-check mess Grafana Labs admits all its codebase are belong to someone who popped its GitHub account Linus Torvalds says AI-powered bug hunters have made Linux security mailing list ‘almost entirely unmanageable’ OpenAI caught in TanStack npm supply chain chaos after employee devices compromised MPs want social media treated more like unsafe toys than harmless apps Nobody believes the 'criminals and scumbags' who hacked Canvas really deleted stolen student data Cops arrest man suspected of being Dream Market kingpin Dirty Frag gets a sequel as Fragnesia hands Linux attackers root-level access To gain root access at this company, all an intruder had to do was ask nicely To gain root access at this company, all an intruder had to do was ask nicely AI models are getting better at replacing cybersecurity pros on certain tasks Cisco to fire 4,000 staff and generously give them free training – on Cisco Welcome to the vulnpocalypse, as vendors use AI to find bugs and patches multiply like rabbits AWS patched Quick auth bypass, says customers weren't using control AWS to Quick admins: The access control didn't work, but you weren't using it anyway, so what's the problem? Bug hunter tracks down three massive MCP flaws and one vendor won't fix theirs Disgruntled researcher releases two more Microsoft zero-days Malware crew TeamPCP open-sources its Shai-Hulud worm on GitHub Vietnam to develop domestic cloud so it can ditch risky overseas operators for government workloads Doozy of a Patch Tuesday includes 30 critical Microsoft CVEs Foxconn confirms cyberattack after Nitrogen claims Apple, Nvidia data theft US bank reports itself after AI customer data mishap Cache-poisoning caper turns TanStack npm packages toxic Apple, Google drag cross-platform texting into the encrypted age Japan’s PM orders cybersecurity review to stop Mythos going full CyberZilla Double Canvas breach acknowledged as ShinyHunters sets new pay-or-leak deadline Cookie thieves caught stealing dev secrets via fake Claude Code installers
微软开源了代理式AI安全工具
2026-05-21 · via The Register - Security

注册

安全

微软冲击 RAMPART,为代理式 AI 安全增添清晰度

雷德蒙德开源两款用于构建和维护更安全代理的工具

微软周三开源了两款旨在帮助开发者和安全团队构建和维护更安全 AI 代理的工具

第一款被称为RAMPART,意为风险评估和测量 用于自主红队演练的平台。这是一个基于微软开源技术的用于自主人工智能应用的pytest框架。PyRIT一个将自动化红队测试嵌入到CI/CD管道中的工具包。 

这允许开发人员模拟现实世界的攻击场景——例如提示注入——并验证代理始终在批准的工具使用、操作和行为边界内。它还支持统计试验,这意味着团队可以设定政策,例如“此操作必须在至少80%的运行中是安全的”,以考虑模型的概率行为.

REG AD

此外,它允许红队和事件响应人员重现任何AI安全发现,以确保代理按预期行为——并且安全缓解措施按预期工作.

REG AD

“是时候我们停止将AI安全视为一种哲学,并开始将其视为一个工程学科了,”微软的数据牛仔拉姆·辛哈·西瓦·库马尔和创始人告诉The Register, 

微软一直在内部使用RAMPART,尽管库马尔表示他无法提供具体细节,但他告诉我们,一名安全研究员发现了一个问题,然后雷德蒙德的红队使用RAMPART来测试该漏洞在智能代理AI应用程序中的存在。

“RAMPART能够识别出那个特定的向量,并找到近100种该向量的变体,”库马尔说。“然后我们利用RAMPART对这一资产进行检测,验证其有效性,不是一次,也不是两次,而是近300次。我们还能在多轮对话的背景下进行检测。”

测试框架还允许开发者在产品中构建缓解措施。 

“他们再次能够使用RAMPART来验证那种补救措施是否有效,不仅针对安全研究员发现的一个攻击向量,还针对这些攻击向量的多种变体,”库马尔解释道。“这增强了我们的事件响应人员和工程师的能力。”

微软周三开源的第二个人工智能工具是一个名为Clarity的代理程序。,根据 Kumar 在周三撰写的关于这两个新工具的博客blog,它旨在作为一个“结构化的意见箱,帮助团队在编写一行代码之前确定他们是否正在构建正确的东西。”

例如,假设一个开发者想要为一个文档编辑器添加实时协作功能。他们向Clarity提出这个需求,代理随后提出的问题类似于“经验丰富的架构师、产品经理和安全工程师会问的问题”,微软表示。

Clarity 的答案,如在 GitHub 上的截图所示:“在我们设计那个之前 - 当两个人同时编辑同一个段落时会发生什么?你需要真正的实时(光标、存在感),还是‘没有人会丢失工作’才是实际要求?这会导致非常不同的架构。”

REG AD

这个AI工具本质上旨在回答开发者试图用应用程序解决什么问题,什么可能会出错,并在编码开始之前“讨论”这些问题。 

“这本质上就是协作,”库马尔说。“它帮助团队退一步思考,‘嘿,在构建这个之前,我们是不是走对了方向?因为代码很便宜。用一根手指的功夫就能生成一个完整的系统。我们是不是在以一种合理的方式来做这件事?’” ®