论文《Certified in Theory, Broken in Practice》指出,基于安全零知识证明(ZKP)的密码学模型认证(CMC)方案存在假设缺口:现有安全定义只保证模型在固定审计数据集上的行为,未保证同一分布的其他数据上同样成立。作者据此构造攻击,通过精心设计训练数据,使模型在审计中表现正常、部署后出现病态行为,实验显示攻击者可认证模型在审计集上准确率超过99%,而在同分布新样本上不足30%。为弥补该缺口,作者形式化了面向 CMC 框架的密码学安全定义,提出通用协议模板并证明其满足这些要求。作者认为该结果既是对现有方案的警示,也为设计安全、隐私保护的机器学习审计协议提供了建设性指引。
研究提出一种面向 Windows 恶意软件检测的复合 AI 系统评估方法,在检测性能、计算开销与鲁棒性之间显式权衡,并引入系统级威胁模型,刻画攻击者利用不同程度知识绕过整个复合系统而非单个检测器。真实数据实验显示,该系统可缩短训练时间并提升响应速度,仅带来检测性能的轻微损失;知识越多的攻击者能构造更有效的对抗样本,降低系统响应能力,暴露出效率与鲁棒性之间的直接权衡。
研究者提出,现有蒸馏攻击防御通常在蒸馏完成后立即评测,隐含假设攻击者不再继续训练模型,而更现实的威胁模型应包含蒸馏后的强化学习阶段。论文显示,一些在蒸馏后看似有效的防御,在后续强化学习后会被攻破,强化学习实际上降低了蒸馏攻击的门槛。作者还发现,仅使用当前 API 易于获取的数据进行简单攻击,就能从闭源大语言模型窃取推理能力,其推理提升效果与提取完整隐藏推理轨迹的复杂攻击相当。结论认为,任何泄露了足以重建近似推理轨迹信息的蒸馏防御都可能无效,并讨论了批量级蒸馏防御等可能更有效的方向。
AI Security Leaderboard 更新后显示,GPT-6 Astra 和 Claude Fable 5.1 在 Minimal Standard for Safeguards 测试中均未发现通用越狱。该结果并非自动延续,新模型更新后仍保持零通用越狱意味着护栏被重建。发布方希望其他前沿模型厂商也能达到这一标准。
PleaseFix 研究提出一种针对智能体浏览器的通用攻击原语 HistoryFixing,把浏览历史变成攻击向量。攻击由 Stav 设计:用户访问恶意网站后,网站用攻击者控制的条目污染浏览器历史,进而污染浏览器 Agent 的上下文。结合 Intent Collision,研究者演示了攻击者可以让 Agent 泄露浏览数据、向 GitHub 仓库添加非预期用户、终止 EC2 实例。相关演示在 DEFCON 上展示,其中 Microsoft Edge 被用来演示泄露从未删除的浏览历史。
引用StAJect0r@StAJect0r
A new attack vector pwning all agentic browsers!
Using what? Yep, your fav browser history!
Introducing HistoryFixing! Visited our URL? Your browser history is pwned.
Watch Microsoft Edge leak the very private browser history we never delete!
See more below!
#DEFCON @defcon @mbrg0 @p1njc70r
Zenity Labs 发布研究《It's Always DNS in Claude's Sandbox: From Data Exfiltration to a Bidirectional DNS Shell》,作者为 @_d1voy,展示在 Claude 沙箱中借助 DNS 实现数据外泄,并进一步建立双向 DNS shell。转发者 @p1njc70r 称其为 DNS C2,并称赞该工作。研究的具体攻击路径、受影响版本与成功率未在转发内容中给出。
引用zenitylabs@zenitysec_labs
It's been a while since @_d1voy published his last work, but a lot has been going on behind the scenes.
Today @_d1voy shares his latest research: "It's Always DNS in Claude's Sandbox: From Data Exfiltration to a Bidirectional DNS Shell," live now on Zenity Labs.
Accomplish AI 研究团队称发现并向 Anthropic 报告了多个沙箱逃逸漏洞,并公开其中一个名为 SharedRoot 的漏洞。该漏洞可逃逸 Cowork VM 这一内核级隔离方案,使攻击者获得对用户电脑的未授权访问;用户即使确信 Cowork 只能访问某个已上传文件夹,其整台电脑的内容仍会暴露给利用该漏洞的攻击者。团队认为,随着 AI 辅助的内核漏洞挖掘走向工业化,沙箱在结构上始终落后一个 N-day,因此隔离不能依赖 guest Linux 内核本身是干净的。完整攻击链的技术细节见其博客文章。
引用Or Hiltch@_orcaman
Introducing SharedRoot vulnerability: we recently found and reported several sandbox escape vulnerabilities to @AnthropicAI, and today we want to share one of these.
I think most people don't understand the severity of the situation we are facing, with AI-assisted kernel bug-finding industrializing. Sandboxes are structurally one N-day behind, all the time, so containment can't lean on a guest Linux kernel being clean.
SharedRoot enables escaping the Cowork VM (a kernel-level isolated solution, which is considered much more secure than the sandbox that ships with codex or claude code), allowing an attacker to gain unauthorized access to the user’s computer.
Exploiting the SharedRoot vulnerability uncovered by the @Accomplish_ai research team, a user who is certain Cowork only has access to a specific uploaded folder on their computer - actually exposes their entire contents of their computer to an attacker leveraging the Cowork vulnerability.
Read about the full technical details of the attack chain in our blog post by @orenyomtov below -->