Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Anthropic launches Anvil automated red-team system

AI summary

Anthropic launched Anvil, a Claude-based automated red-team system that searches for weaknesses in model safeguards across biosecurity, cybersecurity, and information-security tests. Anthropic reported that Anvil found more severe jailbreaks than humans in 23 of 24 biosecurity test combinations, breached all 24 cybersecurity task-model combinations in 0.5 to 3.5 hours, and found its first substantive reasoning-information leak after 435 requests. These benchmark results were not independently verified in the supplied report.

In-depth insight

Anthropic says Anvil could make pre-deployment safeguard testing faster and broaden its coverage. Its tests also found that using internal Claude models with safeguards disabled was important, while precise classifier scores were not decisive because Anvil could still make progress with interception-only feedback.

智东西 编译 | 黄菲 编辑 | 江宇

智东西11月13日消息,今日,Anthropic公布了一套名为Anvil的自动化红队测试系统。该系统基于Claude构建,能够自主寻找模型安全防护机制中的漏洞,并在生物安全、网络安全和信息安全三个领域展开测试。Anthropic称,Anvil在多项测试中的表现已经超过人类红队测试人员,部分漏洞发现速度从数周缩短至数小时。

Anthropic新AI红客来了!只用人类1/235的请求量,就找到严重漏洞

▲Anthropic官方技术博文。来源:Anthropic

随后,Anthropic安全与AI安全研究员Javier Rando也在X平台发文介绍了这项工作,并感谢了Anthropic官方技术文章列出的前三位共同作者Nate Rahn、Chawin Sitawarin和David Huang的贡献。

#Anthropic#Anvil#Claude#Investigator Agent#ExploitBench V8#ExploitGym#Cybertasks#Opus 5.5#Fable 5.1

Original · 13 min readRead the original
Group chat

No comments yet. Start the conversation.