
IT之家 10 月 31 日消息,OpenAI 于当地时间 10 月 30 日发布博文表示,维护社区安全始终是 OpenAI 的重中之重。其日常工作涵盖主动向执法机关通报迫在眉睫的暴力威胁。仅在 2026 年第三季度,OpenAI 就向执法部门移交了 200 余起案件线索。
OpenAI 承诺定期发布《社区安全报告》(Community Safety Report),平台公开在暴力预防、其他社区安全议题以及各项防护机制实际成效维度的进展。首期报告将于 2027 年初正式发布,涵盖 2026 年下半年的安全数据,此后每半年更新一次。
IT之家获悉,OpenAI 针对 ChatGPT 如何回答有关既往大规模恶性袭击事件及凶手的难题增设了额外限制:即便相关信息已在网络公开发布,平台也会严格受控,仅给出极度精简的信息。与此同时,公众仍然可以借助 ChatGPT 了解此类事件的全貌、受害者与社区受到的影响,并学习如何识别危险预警信号以防范将来暴力。
但 OpenAI 进一步收紧了安全策略,坚决限制对既往大规模袭击事件的细节复盘,涵盖所涉武器装备的参数、获取途径,以及可能助长凶手恶名的伤害规模比较。无论使用者的年龄大小或提问动机如何,上述限制均无差别适用。
另外,当对话表明使用者开始执念于暴力、或萌生严重伤害他人的念头时,ChatGPT 可自动终止当前对话,从而即时阻断这种心理与行为倾向。若同一账户因该原因导致对话被多次强制中断,OpenAI 可能会临时冻结该账户。
OpenAI 还宣布,对青少年可获取的涉枪涉武器信息增设了更严格的门槛,即便相关提问表面看似无害也会受到限制。即便使用者未显露任何暴力意图,平台也将全面限制给出可能被滥用以滋生暴力的武器实用知识以及安防平台绕过信息。
OpenAI 当下正加紧开发新能力,以便在青少年账户因反复讨论暴力内容被临时冻结时,第一时间向家长推送通知。
本文来源:https://www.ithome.com/1/020/394.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to IT之家 reader South China Wu Yanzu for the tip!
According to IT之家 on October 31, OpenAI announced in a blog post on October 30 local time that maintaining community safety has always been a top priority for OpenAI. Its daily work includes proactively reporting imminent violence threats to law enforcement. In Q3 2026 alone, OpenAI referred over 200 case clues to law enforcement.
OpenAI commits to periodically publishing the Community Safety Report, in which the platform discloses progress in violence prevention, other community safety issues, and the actual effectiveness of various protection mechanisms. The first report will be officially released in early 2027, covering safety data from the second half of 2026, and will be updated biannually thereafter.
IT之家 has learned that OpenAI has added extra restrictions on how ChatGPT answers questions regarding past mass violent attacks and perpetrators: even if the relevant information has been publicly released online, responses will be strictly controlled to only provide extremely concise information. Meanwhile, the public can still use ChatGPT to understand the overall context of such incidents, the impact on victims and communities, and learn how to identify danger warning signs to prevent future violence.
However, OpenAI has further tightened its safety policy, firmly restricting detailed breakdowns of past mass attacks, which include the specifications and acquisition channels of the involved weapons, as well as casualty comparisons that could boost the perpetrator's infamy. These restrictions apply uniformly regardless of the user's age or intent behind the query.
Additionally, when a conversation indicates that a user is becoming obsessed with violence or contemplating serious harm to others, ChatGPT can automatically terminate the current conversation, thereby immediately interrupting such psychological and behavioral tendencies. If the same account has its conversation forcibly terminated multiple times for this reason, OpenAI may temporarily freeze the account.
OpenAI also announced stricter thresholds for firearm and weapons information accessible to minors, restricting even seemingly harmless queries. Even if a user shows no violent intent, the platform will comprehensively restrict providing practical weapon knowledge and security platform bypass information that could be misused to incite violence.
OpenAI is currently accelerating the development of new capabilities to immediately notify parents when a minor's account is temporarily frozen due to repeated discussions of violent content.