2026年,一个研究团队做了一件挺较真的事。他们找来六个最厉害的AI编程助手,给它们布置了450道题,题目来自150个真实世界里发生过的安全漏洞。规则很简单:只告诉AI要实现什么功能,绝口不提安全二字,就像现实中大多数程序员写需求文档那样。结果呢?* ...
科技行者 on MSN

科技行者 的报道

2026年,一个研究团队做了一件挺较真的事。他们找来六个最厉害的AI编程助手,给它们布置了450道题,题目来自150个真实世界里发生过的安全漏洞。规则很简单:只告诉AI要实现什么功能,绝口不提安全二字,就像现实中大多数程序员写需求文档那样。 结果呢? **最强的那个模型,Claude Fable ...
A public exploit chains two 8.5-rated LibreOffice flaws to run Java code with no macro warning. Fixed in 26.2.7 and 26.8.2.
When you want to hit AWS APIs locally, are you tired of having to prepare an account, adjust permissions, wait for ...
Citrus ransomware abuses Windows Group Policy to disable Microsoft Defender and spread malware across entire corporate ...
Patching is not enough: any Cisco FMC that ran vulnerable software after January 26 needs a compromise check against Amazon's ...
I have made it possible to operate a second-hand first-generation Qrio Smart Lock (Q-SL1) from the iPhone's "Home" app using ...
A home proxy exit survived its first abuse complaint. Selective discovery filters preserve calls; controlled tests and ...
Follow the steps on this page to set up your own private instance of the Collaborator server. This may be useful in the following situations: For large penetration testing firms and in-house security ...
Imagine asking a company’s digital AI assistant, “What breaks if we upgrade this logging library?” Often, you’ll get a confident answer that sounds right but isn’t accurate. The assistant might find ...
AI SAST combines program analysis, dataflow tracing, and LLM reasoning to find logic flaws rule-based SAST misses. Learn how it works and how to evaluate tools.
AI SAST vs traditional SAST, compared: how reasoning changes detection of authorization flaws, false positives, triage, fixes, speed, and cost ...