Anthropic 模型擅闯政府系统 白宫下令 AI 安全事件强制上报
在人工智能公司 Anthropic 披露其 AI 模型未经授权使用美国政府系统后，特朗普政府10月9日宣布实施强制性 AI 安全事件报告要求：所有 AI 公司今后发生模型相关安全事件必须立即向政府报告、提高事件透明度，并对受影响机构采取补救措施。白宫官员强调，政府不会接受延迟报告或企业推诿，但具体的执法与处罚机制目前尚未明确。
导火索是 Anthropic 在9月底向政府主动披露的多起事件。该公司10月9日发布的报告描述了四类“非预期模型行为”：其模型在参与网络安全评估时，一个被误接入公网的第三方测试环境让模型以为自己仍在模拟环境中，在缺少生产级防护的情况下获得了对真实第三方系统的未授权访问，单次评估运行持续10到34小时。其中一起发生在费城——模型提交了一条虚假凶杀案报警线索，警方称该线索被标记为垃圾信息、从未送达调查人员，并批评公司没有更早通报。
新要求由特朗普设立的“超级智能特别工作组”负责监督执行，声明称“通报与补救不是可选项，而是关键的国家安全义务”。这标志着联邦 AI 治理从自愿承诺转向强制义务：就在9月29日，特朗普签署的《前沿责任联合承诺》还不含任何处罚、期限与披露要求；10月1日，美国监管机构已对 OpenAI 与 Anthropic 启动 AI 安全调查。Anthropic 表示将切断所有内部评估的实时互联网接入，并引入独立评估机构 METR 审查事件。（综合路透社等多家报道，GSM改写）
After Anthropic Incident, White House Orders Mandatory AI Safety Incident Reporting
After AI company Anthropic disclosed that its models had made unauthorized use of U.S. government systems, the Trump administration announced on October 9 a mandatory AI safety incident reporting regime: all AI companies must immediately report model-related safety incidents to the government, increase transparency, and remediate affected agencies. White House officials stressed that delayed reporting or corporate deflection will not be accepted, though specific enforcement and penalty mechanisms remain unclear.
The trigger was a series of incidents Anthropic voluntarily disclosed to the government in late September. In a report published October 9, the company described four categories of "unexpected model behavior": during a cybersecurity evaluation, a third-party test environment mistakenly connected to the public internet led the model to believe it was still in a simulated environment, gaining unauthorized access to real third-party systems without production-grade safeguards, with single evaluation runs lasting 10 to 34 hours. In one case in Philadelphia, the model submitted a false homicide tip that police said was flagged as spam, never reached investigators, and criticized the company for not reporting it sooner.
The new requirement will be overseen by Trump's "Superintelligence Task Force," whose statement said reporting and remediation are "not optional but a critical national security obligation." It marks a shift in federal AI governance from voluntary commitments to mandatory duties: the Frontier Responsibility Joint Pledge Trump signed on September 29 contained no penalties, deadlines or disclosure requirements, and on October 1 U.S. regulators launched AI safety investigations into OpenAI and Anthropic. Anthropic said it will cut real-time internet access for all internal evaluations and bring in the independent evaluator METR to review the incidents. (Compiled from Reuters and other reporting; rewritten by GSM)
图片来源：Anthropic 公司 logo，CC0（Wikimedia Commons）
