“人类历史从未让我们有过面对比我们更聪明的行为体的经验。它能够超越我们的思考。”英国国内安全局（MI5）局长肯·麦卡勒姆爵士在10月13日谈到人工智能时如此表示。他在MI5伦敦总部发表年度威胁评估报告。在同一次演讲中，他宣布成立一个新的人工智能威胁评估中心，即AITAC。
“Human history has given us no experience of facing an actor which is smarter than we are. Is capable of out-thinking us.”Sir Ken McCallum, director general of MI5, said this about AI on 13 October. He was giving his annual threat update at MI5’s London headquarters. In the same speech, he announced a new AI Threat Assessment Centre, known as AITAC.
他还表示，MI5现在用于应对国家威胁的资源比冷战时期更多。其国家威胁案件数量在过去十年里增加了四倍。“区分虚张声势与现实”的中心AITAC设在MI5内部。它汇集了来自人工智能安全研究所、国家犯罪局和国家网络安全中心的专家。麦卡勒姆表示，其任务是“区分虚张声势与现实”。这是MI5的第三个此类中心，之前已有针对恐怖主义和国家威胁的部门。
He also said MI5 now devotes more resources to countering state threats than it did during the Cold War. Its state-threat caseload has quadrupled over the past decade, he said. A centre to ‘separate the bluster’AITAC sits inside MI5. It draws on experts from the AI Security Institute, the National Crime Agency and the National Cyber Security Centre. Its job is to “separate the bluster from the realities”, McCallum said.
被问及其运作方式时，他表示，这将为政府提供“一个统一的真相版本”。《欧洲政客》网站如此报道。没有它，不同部门可能会提出相互竞争的观点。
It is MI5’s third centre of this kind, after its units on terrorism and state threats. Asked how it would work, he said it would give government “a single version of the truth”, Politico Europe reported. Without it, different departments might offer competing views.
麦卡勒姆说，人工智能已经在“加强”网络攻击者与防御者之间的竞争。它还能够生成针对每个人的定制宣传。
McCallum said AI is already “supercharging” the contest between cyber attackers and defenders. It can also produce propaganda tailored to each person.
“我们已经抓获了一些试图利用人工智能来改进侦察和攻击方法的恐怖分子，”他说。他补充说，具有人工智能能力的无人机带来了“新的变化”。
“We’ve caught terrorists seeking to exploit AI to improve their reconnaissance and attack methodologies,” he said. He added that AI-enabled drones bring “a new twist”.
他最担心的滥用尚未发生。那是人工智能有可能改变现有甚至全新生物武器的风险程度，他说。生物武器风险有多近？
The misuse he worries about most has not happened yet. It is AI’s potential to change the stakes on existing or even new bioweapons, he said. How close is the bioweapon risk?
《卫报》科学编辑伊恩·桑普尔向专家询问了这一风险的真实性。伦敦国王学院的菲利帕·伦特佐斯博士在接受《卫报》解释性报道采访时表示，更紧迫的担忧是资源充足的团体和国家。
Ian Sample, the Guardian’s science editor, asked experts how real that is. Dr Filippa Lentzos of King’s College London told the Guardian’s explainer that the more immediate concern is well-resourced groups and states.
“在这里，人工智能并不消除对专业知识的需求，而是放大已经存在的专业知识，”伦特佐斯说。
“Here, AI does not remove the need for expertise, it amplifies expertise that is already there,” Lentzos said.
她说，信息只是问题的一部分。专业知识、材料、设备和交付仍然是巨大的障碍。约翰斯·霍普金斯大学健康安全中心的汤姆·英格尔斯比教授表示，目前还没有哪个国家针对这种滥用行为制定正式政策。
She said information is only one part of the problem. Know-how, materials, equipment and delivery remain big barriers. Prof Tom Inglesby, of the Johns Hopkins Center for Health Security, said no country has formal policies yet against this kind of misuse.
给初创企业的信息 麦卡勒姆还将国家韧性（MI5称之为“保护”）列为与恐怖主义和国家威胁并列的第三项核心使命。他最后向包括创始人在内的不同群体提出了建议。
A message for startups McCallum also made national resilience, which MI5 calls “Protect”, a third core mission alongside terrorism and state threats. He closed with advice for different groups, including founders.
“如果你是一家科技初创公司，要预料到会受到外国行为体的间谍活动，并做好网络安全工作，”他说。
“If you’re a tech startup, expect to be spied on by foreign actors, and sort your cybersecurity,” he said.
两周前，MI5曾向英国大学发出间谍活动警报。该警报涉及一个资助人工智能和网络安全研究的研究所。
Two weeks earlier, MI5 had issued an espionage alert to UK universities. It concerned an institute funding research on AI and cybersecurity.
OpenAI无法排除更多事件 麦卡勒姆的演讲与下议院的一次听证会同时举行，据《法庭新闻社》报道。商业、创新、科学和贸易委员会询问了OpenAI、Anthropic、谷歌DeepMind和Meta的高管。议员们询问自愿性保障措施是否能跟上更强大的人工智能的步伐，以及关于人工智能代理以意想不到的方式访问政府网站的问题。
OpenAI cannot rule out more incidents McCallum’s speech coincided with a hearing in the House of Commons, Courthouse News reported. The Business, Innovation, Science and Trade Committee questioned executives from OpenAI, Anthropic, Google DeepMind and Meta. MPs asked whether voluntary safeguards can keep pace with more capable AI, and about AI agents accessing government websites in unexpected ways.
OpenAI负责欧洲、中东和非洲政策的汤姆·达夫·戈登无法说那些就是全部事件。“我不能向委员会或您，主席，保证这就是全部，”他说。
Tom Duff Gordon, OpenAI’s head of policy for Europe, the Middle East and Africa, could not say those were the only ones. “I cannot assure the committee or you, Chair, that that is it,” he said.
他说，他知道还有“少数几起”OpenAI已报告的事件。在七月份的Hugging Face事件之后，OpenAI暂停了对其最先进模型的强化学习训练，他说。它还推迟了一次发布，即Astra 6.1。
He said there were “a handful of others” he knew of that OpenAI had reported. After the Hugging Face incident in July, OpenAI paused reinforcement-learning training on its most advanced model, he said. It also held back a release, Astra 6.1.
前研究人员反驳 前实验室研究人员持批评态度。现任AI Futures Project负责人的丹尼尔·科科塔约洛解释了他离职的原因。
Former researchers push back Former lab researchers were critical. Daniel Kokotajlo, who now leads the AI Futures Project, explained why he quit.
“我离开OpenAI，是因为我开始相信OpenAI实际上不会在创造这种技术方面表现负责任，”科科塔约洛说。
“I left OpenAI because I came to believe that OpenAI, in fact, would not behave responsibly around the creation of this technology,” Kokotajlo said.
曾在OpenAI从事产品安全工作的史蒂文·阿德勒表示，公司内部“对使用‘危险’等词汇存在抵触”。
Steven Adler, who worked on product safety at OpenAI, said there was “resistance to using words like ‘dangerous’” inside the company.
Geoffrey Irving 在 2020 年离开 Google DeepMind，被问及他关于 AI 有 50% 概率毁灭人类的估计。他说抛硬币是“大约正确的模型”。
Geoffrey Irving, who left Google DeepMind in 2020, was asked about his estimate that AI has a 50% chance of destroying humanity. He said a coin flip is “roughly the right model to have”.
Anthropic 公司欧洲、中东和非洲地区政策负责人 Lord Booth-Smith 表示，防护措施必须与能力同步提升。他引用了首席执行官关于“前沿发展需同步推进”的呼吁。
Lord Booth-Smith, Anthropic’s head of policy for Europe, the Middle East and Africa, said safeguards must keep pace with capabilities. He pointed to his chief executive’s call for “a pacing of the frontier”.
委员会主席 Liam Byrne 宣布将启动一项关于 AI 安全的新调查。委员会还将在政府可能对《网络安全法案》进行修改之前发布一份特别报告。
Committee chair Liam Byrne announced a new inquiry into AI safety. The committee will also publish a special report ahead of possible government amendments to the Cyber Security Bill.