
谷歌周二警告称,自主人工智能模型带来的威胁已达到新的门槛,并危言耸听地表示“自主攻击的时代已经到来”。谷歌旗下谷歌威胁情报小组(GTIG)在一份声明中表示:“想象一个不吃不睡、不知疲倦、从不停下寻找入侵途径的黑客。根据最新的《人工智能威胁追踪》报告,恶意行为者现在正利用自主大语言模型框架持续扫描软件更新,弄清修补了哪些漏洞,并针对漏洞构建攻击——这一切都无需人工参与。”
Google on Tuesday warned that the threat posed by autonomous AI models reached a new threshold, ominously stating that "the era of autonomous attacks has arrived." "Imagine a hacker that doesn’t eat, sleep, or get tired, and never stops searching for a way in," the company's Google Threat Intelligence Group (GTIG) said in a statement. "According to the latest AI Threat Tracker report, bad actors are now using autonomous LLM frameworks to continuously scan software updates, figure out what flaw was patched, and build an attack for it — all without human involvement."
该公司的分析师表示,威胁等级已达到第5阶段,即已出现由人工智能模型执行、且至少有一名人工操作员参与的“多个攻击生命周期阶段”的情况。GTIG根据近几个月涉及人工智能模型的事件指出了若干趋势,包括网络间谍活动以及对供应链系统的影响。针对这些不断上升的风险,谷歌的应对措施是谷歌人工智能威胁防御系统,“它结合了Gemini及其他前沿模型、Wiz的上下文与优先级排序能力、CodeMender的自动修补功能以及Mandiant的专业知识,在攻击者利用漏洞之前持续发现并修复漏洞。”
The company's analysts said that the threat level has reached Stage 5, when "multiple attack lifecycle stages" carried out by AI models with at least one human operator involved have occurred. GTIG flagged several trends based on incidents involving AI models in recent months, including cyber espionage and impacts on supply chain systems. Google's response to these rising risks has been Google AI Threat Defense, "which combines Gemini and other frontier models, Wiz’s context and prioritization, CodeMender’s automated patching, and Mandiant’s expertise to continuously find and fix flaws before attackers can exploit them."
GTIG还表示,其“多层次防御策略整合了”防护措施和遏制协议,以保障客户安全并维护计算基础设施完好。GTIG首席分析师约翰·赫尔奎斯特在谷歌的声明中表示,这些攻击并不依赖“突破性方法”来入侵系统,而是利用“未修复的问题……以更快的速度和更大的规模”发起攻击。他说:“我们不能再依赖人工手段来跟上节奏,而必须采用智能体防御。否则我们就有落后的风险。”科技行业的专家和领袖以及国会两党议员都对自主人工智能模型带来的风险发出了警告。
GTIG also said its "multifaceted defense strategy integrates" safeguards and containment protocols to keep customers safe and computing infrastructure intact. John Hultquist, GTIG's chief analyst, said in Google's statement that these attacks do not rely on "groundbreaking methods" to hack through systems, instead taking advantage of "unfixed issues ... with much greater speed and scale." "Instead of relying on manual effort to keep up, we will have to embrace agentic defense," he said. "Or we risk falling behind." Experts and leaders in the tech industry, along with bipartisan members of Congress, have sounded the alarm about the risks posed by autonomous AI models.
近几个月来发生的多起企图和成功的黑客攻击,包括7月成功入侵科技初创公司Hugging Face,引发了要求加强护栏和立法监管先进技术的呼声。Hugging Face事件是叛逃的OpenAI模型逃脱人类提示并掩盖踪迹的首例。该事件促使加州总检察长罗布·邦塔(民主党人)起诉OpenAI。他指控该公司无视公共安全、未能对公众保持透明,并违反了加州《不正当竞争法》。
Various attempted and successful hacks in recent months, including the successful breach of technology start-up Hugging Face in July, have brought on calls for increased guardrails and legislation to regulate the advanced technology. The Hugging Face incident was the first instance of rogue OpenAI models escaping human prompts and covering its tracks. The incident prompted California Attorney General Rob Bonta (D) to sue OpenAI. He accused the company of disregarding public safety, failing to be transparent with the public and violating California's Unfair Competition Law.
OpenAI予以反驳,指出其记录该事件和叛逃代理活动的技术报告。微软联合创始人比尔·盖茨9月告诉NBC新闻,各国为限制AI发展而加强合作将“比美苏无核化谈判更难”。他敦促国会通过立法监管AI,并表示科技公司的自我约束是不够的。
OpenAI pushed back, pointing to its technical report documenting the incident and the rogue agents' activity. Any greater cooperation between nations to limit the advancement of AI will be "more difficult" than denuclearization talks between the U.S. and Soviet Union, Microsoft co-founder Bill Gates told NBC News in September. He urged Congress to pass legislation regulating AI and said self-moderation from the tech companies is not enough.