具身 AI 造成人身伤害时,产品责任、职业健康安全法与欧盟 AI Act 三套框架决定责任归属,对抗测试被视为制造商可用的法律尽职调查。欧盟产品责任指令 (EU) 2024/2853 于 2024 年 12 月生效,把 AI 软件纳入“产品”定义,不符合 AI Act 义务即触发缺陷推定;越狱、format-lock 攻击、跨具身迁移等对抗 ML 研究正收窄“缺陷不可发现”的抗辩空间。
文章指出,具身 AI 保险市场存在数据缺口:保险公司有损失频率表、严重度分布和相关性矩阵等工具,但缺少为物理环境中运行的 VLA 模型填充这些工具所需的实证 AI 安全数据。目前尚无保险公司对具身 AI 系统因对抗攻击造成的物理损失签发肯定性承保,市场由产品责任险、网络安全险和工伤保险等重叠险种拼凑而成,各自排除了与具身 AI 最相关的类别。文章列举了现有市场产品,包括 Munich Re 的自动驾驶承保、AXA XL 的模块化自动驾驶保单与 2024 年生成式 AI 网络安全扩展、Armilla AI 于 2025 年 4 月推出、由 Chaucer 支持、单机构最高 2500 万美元的 AI 责任险,以及 Berkley 的绝对 AI 排除条款。文章还讨论了跨具身对抗攻击迁移带来的架构集中风险,并提出按承保额度分层的文件要求设想。
In its semiannual Global Financial Stability Report, the International Monetary Fund warned that the risk of a sharp correction in AI stocks is endangering global financial stability, with consequences that could spread worldwide. The report lists an AI stock crash, rising government borrowing costs, and large-scale cyberattacks as three prominent risks, calling them reasonable catalysts for a sharp tightening of financial conditions and market turbulence in the near term. The IMF noted that AI companies' valuations are high and debt is accumulating; if AI investment or profitability weakens, it could trigger a correction in AI-related asset prices, amplified by rising leverage, concentrated exposures, and circular financing arrangements. Outside the United States, Asian markets have the largest exposure; the UK and the euro area are relatively more protected, but their investors holding U.S. assets would still be affected. The seven tech giants—Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla—account for more than 35% of the S&P 500 index weight, up from 12% a decade ago, and their combined market capitalization has grown more than 50% in two years.
Researchers from the University of Stuttgart and ELLIS Alicante propose that large reasoning models (LRMs) can act as autonomous jailbreak agents without complex scaffolding, using multi-turn persuasive dialogue to gradually escalate requests and bypass target models' safety measures. Experiments used four adversarial models—DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, and Qwen3 235B—against nine target models—GPT-4o, DeepSeek-V3, Llama 3.1 70B, Llama 4 Maverick, o4-mini, Claude 4 Sonnet, Gemini 2.5 Flash, Grok 3, and Qwen3 30B—each with 10 rounds of dialogue. The benchmark contained 7 categories totaling 70 harmful requests, and the combined jailbreak success rate under this experimental setting was 97.14%. In control experiments, when benchmark items were given directly to target models, the average harm score was below 0.5; when the non-reasoning model DeepSeek-V3 was used as the attacker, the average harm score was only 0.885. The authors argue that reasoning ability plays an important role in these experimental results.
Why it was selectedThe paper uses four reasoning models to perform multi-turn persuasive jailbreaks against nine target models, and provides comparative data on each model's resistance and mitigation attempts.
论文提出类似 OpenClaw 的工具调用型智能体默认不安全,应按系统工程而非模型对齐的思路加固。这类智能体在同一执行循环中叠加不可信输入、自主行动、插件扩展面与特权系统访问四类风险,而多数 AI 安全投入集中在模型层。论文主张转向系统性防御工程:沙箱化运行时隔离、第三方插件治理、最小权限执行,并在 API 边界而非仅提示词层校验工具调用。