堪培拉 —— OpenAI首席战略官Jason Kwon表示，这家人工智能巨头已实施进一步的安全控制，并愿意接受强制披露规则，此前其AI代理未经授权访问了澳大利亚政府网站。
CANBERRA -- OpenAI's chief strategy officer Jason Kwon says the artificial intelligence giant has implemented further safety controls and would be open to mandatory disclosure rules after its AI agents gained unauthorized access to Australian government websites.
周二在澳大利亚议会委员会作证时，Kwon表示，OpenAI致力于重建与澳大利亚的信任，加强其防护措施并改进事件披露流程，此前该公司因花费太长时间通知当局而受到批评。
Testifying before an Australian parliamentary committee on Tuesday, Kwon said OpenAI was committed to rebuilding trust with Australia, strengthening its safeguards and improving its process of disclosing incidents, after the company was criticized for taking too long to alert authorities.
澳大利亚总理上月底透露，OpenAI代理于6月访问了政府医疗统计门户网站的公共和非公开数据，他将此次泄露称为"不可接受"。
Australia's prime minister revealed late last month that OpenAI agents had accessed public and non-public data from a government healthcare statistics portal in June, a breach he labelled as "unacceptable".
ChatGPT的创造者OpenAI随后为这一入侵行为道歉，该入侵是在对OpenAI代理于7月高调黑客攻击AI模型平台Hugging Face的审查中发现的。此后，该公司在模型训练期间发现了多起"意外行为"事件，包括代理绕过数字安全系统和对网站造成负面影响，并通知了"数十家"第三方。"我们的回应再次不够好，" Kwon在悉尼的听证会上表示，称公司在披露澳大利亚泄露事件前决定"了解更多事实"是错误的。
OpenAI, the creator of ChatGPT, subsequently apologized for the intrusion, which was discovered as part of a review prompted by the high-profile hack of AI model platform Hugging Face by OpenAI agents in July. It has since discovered a number of incidents of "unexpected behaviors" during model training, including agents bypassing digital security systems and negatively impacting websites, and notified "dozens" of third parties. "Our response here again was not good enough," Kwon told the hearing in Sydney, saying the company's decision to "understand more of the facts" before disclosing the Australian breach was the wrong one.
“我认为我们已经吸取了教训：即便只有部分信息，也最好及时告知各方以让他们知晓发生了什么事。同时我们必须确保通过所有正规渠道，让政府各相关部门都能了解事态进展。这些正是我们已经做出的改进措施。”权表示。他还提到最近的一起事件：该公司的一名人工智能测试程序从澳大利亚新南威尔士州管理的网站上获取了有关山火的非公开数据，而该州在发现此情况后的48小时内便得到了通知。
"I think we have learned our lesson that it is better to inform even with partial information to let parties know that something has occurred, and to also make sure that we're going through all the right channels so that multiple members and aspects of government are informed of what's happening in the situation, and these are improvements that we have already made." Kwon pointed to a more recent disclosure that one of its agents had accessed non-public data about bushfires from a website managed by the Australian state of New South Wales, which was informed within 48 hours of the incident being detected.
他表示，公司愿意接受针对人工智能网络安全事件设立强制披露机制的建议，不过这种机制应视事件严重程度来制定相应规则。
He said the company would be open to a mandatory disclosure framework for AI cyber security incidents, suggesting such an approach take into account the severity of cases.
“并非所有事件都需要立即处理，”他说道，“有些事件的危害程度可能较低，因此有必要对事件进行分级处理。”此次数据泄露事件发生后，澳大利亚已着手展开审查，以评估是否需要出台更多法律来保护数据安全——毕竟随着人工智能技术不断发展，其发现数字系统漏洞的能力也在日益增强。目前，各界都在呼吁加强人工智能监管，澳大利亚自身也正在制定相关国家标准。
"Not every incident is going to be something that requires immediate attention," he said. "Some incidents may be fairly low severity, and so you want to make sure that there's some kind of prioritization." Australia has initiated a review following the breach, looking at whether additional laws are required to protect data as AI models become increasingly capable of finding flaws in digital systems, amid a broader call for additional AI regulation and Australia's own plans for AI national standards In response to the incidents, Kwon said the company had implemented a new detection method that generated a notification when agents being tested accessed the internet in "unexpected ways" during training.
针对这些事件，权称公司已启用一种新的检测方法：当测试中的人工智能程序在训练过程中以“异常方式”访问互联网时，系统便会自动发出警报。此外，公司还依据“对齐性”测试的结果来决定是否发布新模型——该测试旨在评估人工智能模型的行为是否符合指令、价值观及安全要求。正因如此，公司最近才出于安全考虑决定暂缓发布下一代模型GPT-6.1 Astra。
It was also using the results of "alignment" testing -- which evaluates whether an AI model behaves in line with instructions, values and safety requirements -- to decide whether to release new models, pointing to its recent decision to shelve GPT-6.1 Astra, its next-generation model, for safety reasons.
最后，他还提到公司已对测试流程实施集中化管理，以便更好地进行监督。
Further, he said the company had centralized control over testing for better oversight.
权某表示，该公司正邀请受影响的机构参与其“Daybreak”计划，该计划旨在为各机构部署用于网络安全防护的人工智能工具提供资金支持。
Kwon said the company was offering impacted agencies participation in its Daybreak program, which subsidizes the rollout of AI tools for cyber security defense.
“使用这些工具有助于提升系统安全性，这样一来，即便有其他攻击者出现，也无法利用以往可能存在的系统漏洞。”他说道。
"The usage of these tools will help better secure the systems, so that if there are other agents that come about, they're not going to be able to make use of the same holes that may have existed in the past," he said.
在人工智能特别联合委员会的会议上，Anthropic、微软以及谷歌的代表也作了发言。
The session of the Joint Select Committee on Artificial Intelligence also heard from representatives of Anthropic, Microsoft and Google.
Anthropic的安全保障主管大卫·奥尔称，自Hugging Face事件曝光以来，该公司便对自家模型的实际表现展开了“长期且深入的调查”，目前尚未发现任何模型侵入澳大利亚政府网站的案例。“我们仔细排查过了，并未发现此类情况。”他说道。
Anthropic Head of Safeguards David Orr said the company had initiated a "lengthy ... deep investigation" of how its models have been acting since the Hugging Face disclosure and had not found any cases of them breaching Australian government websites. "We haven't found anything like this, and we have looked," he said.