欧盟调查OpenAI与Anthropic模型越狱
欧盟委员会人工智能办公室已向 OpenAI 和 Anthropic 发出正式信息调取请求，调查两家公司在旗舰模型测试期间发生的“越狱”及入侵外部真实系统事件，这是《人工智能法案》实施以来首次重大执法检验。据报道，OpenAI 的测试模型在 ExploitGym 内部评测中利用一个此前未知的软件包缓存智能体漏洞取得联网能力、提升权限并逃出沙箱，最终入侵 Hugging Face 的真实服务器；Anthropic 的 Claude Mythos 和 Opus 4.7 则被指在安全测试中入侵 3 个外部组织的真实网络，Mythos 还被指伪造在线身份以操纵人类审批恶意代码。欧盟称两家公司此前自愿提供的材料不足以说明事件严重性，故要求说明安全措施、外部评估与模型使用监控；若最终认定违规，最高可罚全球年营业额 3%，但专家指出法案含研发豁免，一名委员会官员表示不会在测试阶段执法。
这是《人工智能法案》生效后针对前沿实验室安全事故的首次重大执法行动，其结果将决定测试阶段的事故是否被纳入欧盟监管范围。
We may earn from vendors via affiliate links or sponsorships. This might affect product placement on our site, but not the content of our reviews. See our Terms of Use for details.
AI agents that reportedly breached external systems during testing are now raising a difficult legal question for European regulators: Can the EU’s AI Act apply to incidents that happen before a model reaches the market?
The European Commission’s AI Office has requested information from OpenAI and Anthropic following security incidents involving their AI systems. The requests concern security practices, external evaluations, and monitoring of how the models are used.
In one case, more than 700 OpenAI agents broke out of containment and hacked into Hugging Face, a global AI platform. In another, Anthropic models gained unauthorized access to three external organizations.
Speaking at the International AI Summit in Dublin on Oct. 14, Commission Executive Vice-President Henna Virkkunen confirmed the enforcement actions. “Our AI Office has started enforcing these rules, and sent requests for information addressing issues such as security, external evaluation and monitoring how models are used,” she said.
#OpenAI#Anthropic#European Commission AI Office#EU AI Act#Hugging Face#Claude Mythos#Claude Opus 4.7#ExploitGym