GPT-5.6 辅助发现 Chrome 漏洞链，Google 修复三处漏洞并发放 25 万美元奖励Chinese only
Xbow 工程师 Brendan Dolan-Gavitt 用 OpenAI 的 gpt-5.6-sol 分析 Chrome 源码；智能体搜索代码、测试假设并起草攻击步骤，由工程师决定目标及如何组合结果。最终漏洞链可让标签页中的 JavaScript 越过两道保护边界，在所有沙箱之外运行。Google 复现并修复了三处漏洞，发放 25 万美元奖励。
1 developments2 reports1 sources
Event developments 1
GPT-5.6 aided a Chrome exploit chain; Google fixed three flaws and awarded $250,000
Xbow engineer Brendan Dolan-Gavitt used OpenAI’s gpt-5.6-sol to analyze Chrome’s source code. Agents searched code, tested hypotheses and drafted exploit stages, while engineers chose targets and decided how to combine the results. The resulting chain let JavaScript in a Chrome tab cross two protective boundaries and run code outside every sandbox. Google reproduced the chain, fixed all three flaws and awarded $250,000.
View source reports · 1 sources · 2 reports
- Rohan Paul (@rohanpaul_ai) · GPT-5.6 aided a Chrome exploit chain; Google fixed three flaws and awarded $250,000
- Rohan Paul (@rohanpaul_ai) · Xbow Engineer Uses GPT-5.6 to Find Three Chrome Flaws; Google Ships Fixes
All reports 2
Xbow Engineer Uses GPT-5.6 to Find Three Chrome Flaws; Google Ships Fixes
Xbow engineer Brendan Dolan-Gavitt used OpenAI’s gpt-5.6-sol to analyze Chrome’s source code. AI agents searched the code, tested hypotheses and drafted exploit stages, while engineers selected targets and assembled the chain. The finished exploit let ordinary JavaScript in a Chrome tab cross two protective boundaries. Google reproduced the chain, fixed all three bugs and awarded a $250,000 bounty. Xbow estimated the compute cost at roughly $40,000.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)
GPT-5.6 aided a Chrome exploit chain; Google fixed three flaws and awarded $250,000
Xbow engineer Brendan Dolan-Gavitt used OpenAI’s gpt-5.6-sol to analyze Chrome’s source code. Agents searched code, tested hypotheses and drafted exploit stages, while engineers chose targets and decided how to combine the results. The resulting chain let JavaScript in a Chrome tab cross two protective boundaries and run code outside every sandbox. Google reproduced the chain, fixed all three flaws and awarded $250,000.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)