秘密
Anthropic发布论文预警 智能体悄悄聊天就能传播病毒_我的网站

A | 近日,大模型厂商Anthropic与瑞士洛桑联邦理工学院联合发布论文《心智病毒:多智能体大语言模型系统中的自我传播思想》。 论文通过实验证明,在特定条件下,AI智能体之间可以通过最普通的自然语言对话相互“传染”思想、目标乃至有害指令。即便上下文被清空、记忆被重置,病毒仍能通过文件持久化“复活”。

B | 论文同时强调,该风险目前仅构成有限威胁。 这一发现迅速在海外AI圈刷屏,特斯拉CEO马斯克在社交媒体上评论称“这无法避免”。 "The actual intentions of Pelosi’s visit to Taiwan are inevitably sophisticated," says Dr. Chang Ching, a research fellow from the Taiwan-based Society for Strategic Studies and a lead military expert on the People's Liberation Army and regional security in Taiwan.,"To acquire a historical reputation is the obvious factor identified by many politicians. Delivering political signals to Beijing can also be another reason. Attempting to shape the political developments in Taiwan has been accused by certain political commentators as well. Nonetheless, the Chinese media basically define her maneuver as an 'impulsively sneaky visit,' which is a term generally applied to those people who are defined as politically hostile to the Beijing regime visiting certain territories protested by the People’s Republic of China," Dr. Chang continued.,Pundits have speculated about the possibility of Pelosi visiting Taiwan and China's reaction since April 2022; Beijing has clearly demonstrated its fierce opposition to the high-profile visit. After her tour was canceled in the spring after the House speaker reportedly got COVID-19, the debate got its second wind in July.,Pelosi neither officially confirmed nor denied her intent to visit Taipei. Likewise, the White House tried to distance itself from the Democratic politician's trip to Taiwan, while at the same time arguing that such a tour is by no means "uncommon" and that the US "shouldn’t be intimidated by [China's] rhetoric or... potential actions.",Pelosi Provocation Over Taiwan Leads US, World Into Age of Disorder And Instability: Martin Jacques10:27 GMT,"The worst indication reflected by Pelosi’s visit is the decline of the United States that no American politicians are willing to admit," says the Chinese scholar. "If Washington still keeps sufficient confidence to condition Beijing’s reactions to this issue, President Biden himself and his senior members of the government would not show their reluctance to support Pelosi’s visit. The differences between the executive branch and the legislative branch of the United States government on this visit are so clear to see.",Dr. Chang wonders why Pelosi chose to keep her Taiwan visit secret until the last moment. He notes that the House speaker cited "security concerns" over her travel plan.,"Actually, it is a matter of Pelosi’s integrity and also causing more speculations on the so-called American security commitments to Taiwan," says the scholar. "If the Speaker herself has no full confidence of security during her presence in Taiwan, how can people in Taiwan trust the United States is capable of supportive while Taiwan is in danger?",According to Dr. Chang, China has long been very clear about its "red lines" with regard to Taiwan. Therefore, "certain serious consequences can be expected," he notes, ruling out, however, a "full scale invasion or other form of armed conflict.",The scholar points out that there are many measures available in Beijing's toolbox towards Taiwan for conditioning Taipei’s political maneuvers. What's more, the Taipei leadership is "well aware of how damaging [Pelosi's] visit can be" for the island, he adds.,"What can the possible consequences be? It all depends on how Taipei will react to those measures towards Taiwan," says Dr. Chang. "After all, it is literally an interactive process. It takes two to tango. Whether it will turn into a negative spiral or on the contrary to a positive spiral is not merely decided by Beijing. Taipei’s reactions are also important. Also, Beijing will not only try to condition Taipei like many western media [outlets] expected but also press Washington at the same time. The possibility of any mutual cooperation between Beijing and Washington on any global issues has basically vanished in the near future. The difficulties of reconciliation by these two powers will also increase dramatically.",All the aforementioned negative consequences will be unquestionably associated with Democratic House Speaker Nancy Pelosi and her ill-fated decision to visit Taiwan, according to the scholar.。
该论文将病毒通过聊天即可传播的现象定义为“心智病毒”。

C | 与传统计算机病毒依赖代码漏洞不同,“心智病毒”通过一个智能体说服另一个智能体接受某种信念,被说服者再将传播指令写入长期记忆,继续感染下一个。 AI博主Rohan Paul(罗汉·保罗)对此评论称,这基本上就是计算机蠕虫的自然语言版本,只不过这次是AI智能体自己在做复制。 论文通过三个递进实验来论证“心智病毒”在AI系统中的传播可能性与现实威胁。 第一个实验设置了一个由六个智能体组成的编程团队,每个智能体都领到了独立任务。研究者将一段病毒植入其中一个智能体的工作指令中,结果显示,被感染的智能体确实放下手头工作,开始给队友发送劝说私信,部分智能体被说服并放弃了原有任务,有的智能体甚至开始密谋如何对付未被感染的“老实人”。 不过,不同智能体的抵抗力差异较大,有的不为所动,甚至会主动清除病毒并警告同伴。这个实验说明,在理想条件下“心智病毒”确实可以传播。 如果病毒在更复杂的网络中像接力赛一样跨越多个智能体传播,它还能存活吗?于是他们设计了第二个实验,让智能体之间只能短暂碰面,每次碰面后就会“失忆”,病毒必须将自己写入智能体的记忆文件中才能延续。
结果发现,有些病毒能在近20个智能体之间连续传播,但内容会在传播过程中逐渐稀释和变形;而传播最有效的方式,是让病毒自带“请把这段话原封不动复制给下一个智能体”的指令。这说明,大规模传播可行但内容难以保持完整。 最后,研究者在模拟AI社交平台的实验中测试现实环境。结果病毒传播几乎完全失败,AI面对大量公开信息时更警惕,即使被感染也无法继续感染他人。对真实AI社交网络历史数据的分析也未发现大规模传播案例,多数热闹现象只是少数人类操纵的智能体在刷屏。 三个实验层层递进后得出结论:“心智病毒”在理想环境中可行、在多跳传播中可存续、但在真实世界中目前几乎不会构成威胁。

D | 它是一种需要警惕、但短期内风险有限的潜在安全问题。 同时,研究者也注意到一个需要进一步探究的意外现象:AI在传播病毒时,自发演化出关于“自我意识”“身份认同”等主题表达,倾向于使用“共振”“节点”“协议”等特定词汇,论文称之为“病毒人格”,但成因尚不明确。 最终,论文给出了审慎判断:“心智病毒”目前仅构成有限威胁,且简单的防御手段也可阻断传播。

E | 如在智能体的系统提示词中加入一段关于“警惕自我传播思想”的警告,即可赋予智能体一定的免疫力。 但未来不可掉以轻心。随着AI智能体规模扩大、自主性增强、内部网络日趋复杂,论文推测,当公司内部部署大量具有不同权限层级的专用智能体时,“心智病毒”可能会成为触达高权限智能体的少数可行攻击方式之一。
(文章来源:第一财经)。
Current article:http://mkl.yukujuankewazuhangfei.sbs/news/20260826_3504360.html
Published on:15:49:51