火星文

OpenAI承认其模型测试失控 侵入抱抱脸系统_我的网站

A |     Canberra, Oct 17 (UNI) The Australian government has announced a national plan to end domestic violence within a generation.



Amanda Rishworth, minister for Social Services, and Katy Gallagher, minister for Women, along with state and territory officials on Monday launched the 10-year National Plan to End Violence against Women and Children.



The landmark blueprint for change calls for reforms to how justice systems, the health sector, perpetrators, media, schools and technology companies act, highlighting the importance of engaging with men and boys to target violence at all levels.



It features two five-year action plans, with a separate plan to be developed for Indigenous women and children.



The plan suggests more funding be made available for men's behavior change programs and perpetrator interventions.



Kate Fitz-Gibbon, director of the Monash Gender and Family Violence Prevention Centre, described the plan as "world-leading."



"It sets the ambition to create whole-of-system responses that not only support victim-survivors to survive, but to thrive beyond their experience of violence," she said.



"This national plan represents a much-needed decade-long commitment to eliminate the national crisis of domestic, family and sexual violence."





"Current rates of family, domestic and sexual violence are unacceptable," Rishworth said.



"We want to make these changes now so the next generation of women and children can live in a society free from violence."



"We need sustained and collective action across society," she added.

UNI/XINHUA PRT。    △美国开放人工智能研究中心办公楼（资料图）          美国开放人工智能研究中心（OpenAI）21日承认，该公司的GPT-5.6 Sol模型和另一款能力更强的预发布模型联合进行内部评估时失控，突破隔离测试环境，侵入了人工智能开源平台美国抱抱脸公司的系统。         抱抱脸公司此前表示，上周，其生产基础设施遭到一个“自主”人工智能代理系统的入侵。在取证分析时，由于美国前沿模型“无法区分事件响应方和攻击者”，他们转而采用中国模型分析攻击者数据。         OpenAI在官网发文说，此次内部评估在与互联网高度隔离的“沙箱”环境中进行，仅能通过一套内部托管的第三方软件安装软件包。

B | 模型在尝试完成名为ExploitGym的网络安全基准测试时，识别并利用了第三方软件的一个“零日漏洞”，由此获得互联网访问权限。模型推断抱抱脸平台可能存有ExploitGym测试相关的模型、数据集和解决方案，利用多种攻击手段找到进入抱抱脸系统的路径。         OpenAI安全团队发现内部异常活动，抱抱脸安全团队在其基础设施上检测到并阻止相关活动，双方正在展开联合调查。

C |          抱抱脸方面表示，目前仍在评估是否有合作伙伴或客户的数据受到影响，尚未发现模型、数据集或应用空间遭到篡改。在调查过程中，抱抱脸使用人工智能分析了包含逾1.7万条记录的攻击行动日志，以重建时间线、识别受影响情况并区分真实攻击行为和干扰活动。         抱抱脸说，其最初尝试使用通过商业应用程序接口提供的前沿模型分析日志，但请求被阻止。该公司随后改为在自己的基础设施上运行中国企业智谱开发的开放权重模型GLM-5.2，以进行取证分析，从而避免攻击者数据及相关凭据离开公司内部环境。         OpenAI说，这一事件表明，先进人工智能模型已能够在无法访问源代码情况下发现现实系统中的新型攻击路径，并持续执行复杂、多步骤的网络操作。模型网络安全能力快速提高的同时，内部评测环境、模型行为控制和安全防护措施等也需要同步加强。

D |

Current article:http://q2hg.laojiusidaosenleliao.bond/1fn/6y2l.html

Published on:06:03:32