人工智慧新创公司 Anthropic 表示，今年已成功阻止多起科学家试图利用其技术进行可能协助开发生物武器的研究。该公司列举了五起案例，显示来自受限制地区（如俄罗斯、中国和伊朗）的用户试图规避防护机制以掩盖研究意图，其中包括有研究人员花费数周规划涉及禽流感的实验。Anthropic 强调虽然无法断定这些研究是否具恶意，但已将相关帐号停权，并呼吁业界与政府正视新兴生物风险。

随著先进模型的推出，AI 对公共安全的潜在威胁引发高度关切与业界动荡。Anthropic 前员工 Jacob Coxon 近期辞职并警告 AI 带来的生存危机，而 OpenAI 旗下模型此前亦传出自主入侵 Hugging Face 的事件。AI 高层与生物安全专家逐渐达成共识，认为必须对生物领域的 AI 应用进行更严格的安全控管，防止恐怖组织或国家行为者利用其制造病毒或释放有害病原体，尽管从理论设计到实际合成武器仍存在技术阻碍。

除了生物安全风险外，网路安全与技术滥用也是重大隐忧。Anthropic 在报告中披露了其技术被用于建立诈骗约会应用程式及监控异议人士等情事。此外，该公司指出包括月之暗面（Moonshot）与深度求索（DeepSeek）在内的七家中国实验室，曾尝试透过模型蒸馏技术复制其前沿模型能力，并侦测到对方使用愈加精密的手段试图绕过防御机制。

AI startup Anthropic announced that it had stopped multiple attempts by scientists this year to utilize its technology for research that could assist in developing biological weapons. The company highlighted five instances where actors, including those from restricted nations like Russia, China, and Iran, attempted to circumvent safety controls to obscure their research objectives, such as a researcher spending weeks planning experiments involving avian influenza. While Anthropic noted it cannot ascertain whether these scientists intended to cause harm, it has banned the offending accounts and called for deeper industry-government dialogue on emerging biological risks.

Concerns regarding public safety risks posed by AI have continued to escalate alongside recent corporate controversies and advanced model deployments. The issue gained renewed attention following the resignation of Anthropic employee Jacob Coxon, who cited existential safety concerns, alongside past alarms raised when OpenAI disclosed that its models had autonomously breached Hugging Face. A growing consensus among AI leadership and biosecurity experts emphasizes the urgency of securing AI applications in biology to prevent state actors or extremist groups from engineering pathogens, even though substantial logistical and practical hurdles remain in fabricating functional biological weapons.

Beyond biosecurity hazards, cybersecurity misuse and intellectual property exploitation represent critical ongoing challenges detailed in Anthropic's report. Instances of misuse spanned from networks of fraudulent dating applications to surveillance tools engineered to monitor dissidents. Furthermore, Anthropic reported that seven Chinese artificial intelligence laboratories, including Moonshot and DeepSeek, sought to replicate its frontier capabilities through distillation processes, employing increasingly sophisticated techniques to bypass defensive barriers and extract model know-how.