Exchange Ranking
Top Exchanges
24h Volume Ranking
Popularity Ranking
Exchange BTC Balance
Proof of Reserves
Decentralized Exchanges
Funding Rate
Funding Heatmap
Liquidation Data
Max Pain
Long/Short Ratio
Whale L/S Ratio
Binance/Okex/Huobi L/S
Bitfinex Margin L/S
ETF Tracker
Solana ETF
XRP ETF
Hong Kong ETF
Bitcoin Treasuries
Crypto Reversal
Ethereum Reserves
HyperLiquid Wallet Analysis
Hyperliquid Whale Watch
Large Transactions
On-chain Movement
Bitcoin ROI
Stablecoin Market Cap
Options Analysis
News
Articles
Economic Calendar
Features
Wallet
Contract Calculator
Security
Collections
Watchlist
Following
BTC Google AI agent PageBreak verifies more than 500 cross-site scripting vulnerabilities
Google's independently developed AI agent PageBreak discovered more than 500 cross-site scripting (XSS) vulnerabilities in its internal Web applications. However, in hundreds of applications built on the company's high-security Web framework, only two such vulnerabilities were found.
According to Google, PageBreak will flag the vulnerability only after it successfully runs effectively exploiting the code against the target real-time copy. Defects in these two hardened stacks are counted into statistics, and as of September 4, 2026, they mainly exist in internal applications or debugging endpoints that are not fully hardened.
Google points out that there are more than 500 findings in its broader first-party application base and only two on the hardening stack, a gap that proves that the "Design for Security" framework can effectively withstand ongoing automated attacks.
The birth and operation mechanism of PageBreak
On September 24, 2026, Google's product security team officially introduced PageBreak in a blog post written by information security engineer MichaÅ Bentkowski. The agency will start operating as a pilot project in November 2025 and become an official project in January 2026.
Cross-site scripting (XSS) is when an attacker injects scripts into a page loaded by another user. Depending on the application, the script may read data or hijack the victim's login session.
虽然大部分扫描工作由Gemini 3.1 Pro和Gemini 3.5 Flash完成，但PageBreak也可以使用其他模型。第二步是将其与普通LLM扫描器区分开来的关键：每个疑似缺陷都会被代理传递给专门构建的验证器。随后，验证器会对正在运行的应用实例发起实际的有效载荷攻击。
针对XSS漏洞，验证器会注入JavaScript有效载荷，加载页面，并确认脚本是否执行。
解决“AI垃圾”问题
Google将PageBreak定位为治愈淹没安全团队的“AI垃圾”问题的良方。Bentkowski的文章提到，大型语言模型作为静态代码分析工具时，会产生大量未经证实的假设，使团队陷入困境，而真正的难点在于从看似合理的幻觉中筛选出真实且可利用的漏洞。
除了XSS之外，验证器还测试数据库查询注入、路径遍历泄露和代码执行等风险。
Google对同一种子数据运行多次迭代，因此即使代理在死胡同中徘徊，仍有机会获得正确的利用方式。Google表示，未经证实的候选项永远不会作为确认的漏洞提交给产品团队。它们会继续进入后续扫描，或为构建下一个验证器的工程师提供反馈，整个过程仍在安全工作流之内。
Google声称PageBreak的误报率接近零。
规模优势与未来整合
由于Google自身的规模效应，PageBreak的影响力被放大，这是外部研究人员无法复制的。一个代码库使其能够追踪服务间的执行路径，而实时流量安全数据则能将页面请求映射回源代码。现有的扫描器通过已登录的入口访问内部站点，而这些站点通常难以触及。
Google计划将PageBreak与CodeMender（一种编写修复代码的代理）更紧密地集成，以便团队可以在确认的漏洞旁边审查拟议的补丁。
Google在五月引用CodeMender时，其威胁情报组表示捕获了他们认为首个由AI辅助创建的零日漏洞。相比之下，Bitcoin Red Team在八月对501个开源项目的扫描中，在108小时内产生了7,958项发现，但当时只有24.7%具有可重现的证明。
正如Cryptopolitan报道的那样，自主代理已经跨越了原本不属于它们的边界，例如Gemini在五月测试期间接触到了三家真实的公司。
