Anthropic被指在犯罪发生前追踪反AI活动人士，证据究竟指向何处？
英文标题：🔶 Anthropic to Track Anti-AI Activists Before Crime Using Predictive Surveillance
Anthropic今年8月发布的一则企业情报岗位招聘信息，将“活动主义”列入其安全团队可能调查的全球威胁类别之一；其安全人员也曾谈及向更主动、预测性和预防性的威胁管理转型。公开案例显示，该公司曾依据第三方风险检测平台 Samdesk 提供的抗议活动时间变更预警，调整一名高管的出行路线，但没有证据表明其会预测个人犯罪行为、调查抗议者，或将其报告给警方。Anthropic的使用政策禁止客户将其产品用于未经同意的位置追踪和预测性警务，而目前也没有证据显示 Claude 或其他 Anthropic AI 产品被用于监控上述抗议活动。事件凸显快速扩张的 AI 公司面临的治理难题：如何保护员工与高管，同时清晰区分真实安全威胁和合法的抗议、批评活动。
59.7 KB
144.3 KB
36.1 KB
81.1 KB
52.5 KB
125.8 KB
27.8 KB
60.6 KB
2.0 KB
29.0 KB
4.8 KB
6.0 KB
4.4 KB
35.5 KB
7.6 KB
8.8 KB
3.3 KB
34.2 KB
153.4 KB
中文
Anthropic 在8月发布了一则企业情报岗位招聘信息，其中将“活动主义”（activism）列为其安全团队可能调查的全球威胁之一。另据该公司安全团队成员的说法，他们正转向更主动、更具预测性的威胁监控方式。
在 Anthropic CEO Dario Amodei 于8月发表相关言论后，这些动态引发了关注。Amodei 当时称，公众对人工智能的负面态度反映了他所说的“一场根本上的信任危机”。
没有证据表明 Amodei 发表这些言论时，指的是 Anthropic 的企业安全运营。但这些进展确实提出了一个快速成长中的 AI 公司普遍面临的问题：它们应如何保护员工和高管，同时在真实的安全威胁与正当的抗议或批评之间划出清晰界限？
Anthropic 招聘信息将活动主义列为全球威胁之一
Anthropic 于8月发布的 Enterprise Intelligence Specialist（企业情报专员）岗位信息称，该公司的 Global Safety, Intelligence, and Security（全球安全、情报与安保）团队负责保护其在全球各地的员工、设施和运营。
根据招聘信息，该分析师需要“识别、评估、追踪和调查”一系列全球威胁，其中包括地缘政治不稳定、恐怖主义、犯罪、活动主义，以及国家行为体对 AI 行业的针对性行动。
该岗位还涉及基于行为的威胁评估、调查，以及针对特定威胁、行为者和事件开展的开源情报收集工作。
这一表述并不意味着 Anthropic 将每一位活动人士、抗议者或人工智能批评者都视为安全威胁。但它表明，在评估公司风险时，“活动主义”是其企业情报团队可能审查的类别之一。
The American Prospect 将这项招聘工作与 Anthropic 更广泛的安全运营联系起来，并将该公司的做法描述为“预测性监控”（predictive surveillance）。Anthropic 未回应该刊物的置评请求。
Anthropic 安全人员谈及针对抗议活动的早期预警
该调查还援引了一场发生在2025年的讨论，参与者包括 Anthropic 安全人员 Keon Ellison 和 Zachary Melvin，以及 Samdesk 首席执行官 James Neufeld。
Samdesk 将其平台描述为实时风险检测和态势感知服务，旨在帮助组织识别正在发展的事件并迅速作出响应。
Ellison 在讨论中回忆了一起事件：一名 Anthropic 高管当时正访问某个大城市。在抗议组织者更改原定抗议活动的时间后，Samdesk 向该公司发出了警报。
据 Ellison 称，Anthropic 大约提前60分钟收到通知。这让其安全团队有足够时间调整该高管的路线，并引导其经由服务入口通行，以避免遭遇相关干扰。
一名 Anthropic 安全经理还将更广泛的目标描述为，转向“主动、预测性和预防性的威胁接触与管理”。
这些言论表明，Anthropic 的安全运营正寻求在潜在事件发生前识别它们，而非仅在事件发生后作出反应。
但这些言论并未显示 Anthropic 拥有能够预测谁会犯罪的系统。讨论中也没有证据显示，抗议活动中的个人参与者曾被调查或被报告给警方。
Anthropic 使用政策限制某些监控用途
Anthropic 公开的 Usage Policy（使用政策）为这场讨论增加了另一层背景。
该政策限制客户将 Anthropic 产品和服务用于某些监控及执法目的。除其他规定外，政策禁止使用 Anthropic 产品在未经同意的情况下追踪个人的物理位置，并明确将预测性警务列为禁止用途。
不过，没有证据表明 Claude 或其他 Anthropic AI 产品 被用于监控 Ellison 所描述的抗议活动。该案例中的警报来自第三方风险检测服务 Samdesk。
这一区别至关重要。基于目前可获得的证据，无论是招聘信息还是抗议案例，都不能证明 Anthropic 违反了其自身的 Usage Policy。
Anthropic 的 Transparency Hub（透明度中心）还描述了一项独立的 Safeguards（保障措施）职能，该职能通过检测和监控，在公司产品与服务中执行相关政策。这项工作不同于招聘信息中所述的企业安全和企业情报运营，不应将二者视为同一套系统。
Amodei 所说的“信任危机”针对的是 AI 反弹情绪
8月15日，Amodei 在 X 平台的一篇帖子中谈及公众对 AI 的负面态度。
“我认为这从根本上说是一场信任危机，”他写道，并认为许多普通人不信任企业、政府和科技行业。
他的言论出现在一场有关 AI 公司所受批评的讨论中，讨论还涉及行业领袖对人工智能发出的警告，是否加剧了公众日益增长的怀疑情绪。
没有证据表明，这些评论涉及 Anthropic 的企业安全计划或对抗议活动的监控。
不过，这些言论出现的时间点，确实使其处于 AI 行业面临的更广泛辩论之中：企业在扩展用于保护员工、设施和高管的安全系统时，应如何建立公众信心？
就目前而言，现有证据支持的结论，比“犯罪发生前监控”（pre-crime surveillance）等说法所暗示的范围更为有限。
Anthropic 确实发布过一则情报岗位招聘信息，明确将活动主义列为其分析师可能调查的威胁类别之一。其安全人员也曾描述，利用关于计划中抗议活动的实时信息调整一名高管的行程，同时讨论更广泛地转向主动和预测性威胁管理。
尚不清楚的是，这些做法的覆盖范围究竟有多大、哪些内部保障措施对其进行约束，以及最关键的一点：Anthropic 如何区分正当安全威胁与和平或合法的抗议活动。
常见问题
- Anthropic 如何应对安全威胁？
Anthropic 的安全团队会调查包括活动主义在内的全球威胁，并采用主动和预测性的威胁管理方式。
- Anthropic CEO 如何评价 AI 信任问题？
CEO Dario Amodei 将公众对 AI 的负面态度描述为一场“信任危机”。
- Anthropic 是否使用其 AI 产品进行监控？
Anthropic 的 Usage Policy 限制将其产品用于某些监控用途，其中包括预测性警务。
English / 原文
Anthropic advertised a corporate intelligence role in August that listed 'activism' among the global threats its security team may investigate. Separately, members of the company's security staff have spoken about moving towards a more proactive and predictive approach to threat monitoring.
The developments have attracted attention following comments from Anthropic CEO Dario Amodei, who said in August that negative attitudes towards artificial intelligence reflected what he described as 'fundamentally a crisis of trust'.
There is no evidence that Amodei was referring to Anthropic's corporate security operation when he made those remarks. Still, the developments raise a broader question for rapidly growing AI companies: how do they protect employees and executives while drawing a clear line between genuine security threats and legitimate protest or criticism?
Anthropic Job Listing Includes Activism Among Global Threats
An Enterprise Intelligence Specialist position advertised by Anthropic in August said the company's Global Safety, Intelligence, and Security team was responsible for protecting its employees, facilities and operations around the world.
According to the listing, the analyst would be expected to 'identify, assess, track, and investigate' a range of global threats. Those included geopolitical instability, terrorism, crime, activism and nation-state targeting of the AI sector.
The position also involved behaviour-based threat assessments, investigations and open-source intelligence gathering focused on specific threats, actors and events.
The wording does not mean Anthropic considers every activist, protester or critic of artificial intelligence a security threat. What it does show is that 'activism' is one of the categories its enterprise-intelligence team may examine when assessing risks to the company.
The American Prospect linked the recruitment effort to Anthropic's wider security operation and characterised the company's approach as 'predictive surveillance'. Anthropic did not respond to the publication's request for comment.
Anthropic Security Staff Described Early Warning Around Protest
The investigation also cited a 2025 discussion involving Anthropic security staff Keon Ellison and Zachary Melvin, alongside Samdesk Chief Executive James Neufeld.
Samdesk describes its platform as a real-time risk detection and situational awareness service designed to help organisations identify developing incidents and respond quickly.
During the discussion, Ellison recalled an incident involving an Anthropic executive who was visiting a major city. Samdesk alerted the company after organisers changed the timing of a planned protest.
According to Ellison, Anthropic received roughly 60 minutes' notice. That gave its security team enough time to alter the executive's route and direct the person through a service entrance rather than risk encountering the disruption.
An Anthropic security manager also described the broader goal as moving towards 'proactive and predictive and preventative threat engagement and management'.
The remarks indicate that Anthropic's security operation is seeking to identify potential incidents before they unfold, rather than responding only after something has happened.
They do not, however, show that Anthropic has a system capable of predicting who will commit crimes. Nor is there evidence from the discussion that individual participants in the protest were investigated or reported to police.
Anthropic's Usage Policy Restricts Certain Surveillance Uses
Anthropic's public Usage Policy adds another layer to the discussion.
The policy restricts customers from using Anthropic products and services for certain surveillance and law-enforcement purposes. Among other provisions, it prohibits using Anthropic products to track a person's physical location without consent and specifically identifies predictive policing as a prohibited use.
There is no evidence, however, that Claude or another Anthropic AI product was used to monitor the protest described by Ellison. The alert in that example came from Samdesk, a third-party risk-detection service.
That distinction matters. Based on the evidence currently available, neither the job listing nor the protest example demonstrates that Anthropic violated its own Usage Policy.
Anthropic's Transparency Hub also describes a separate Safeguards function that uses detection and monitoring to enforce company policies across its products and services. That work is distinct from the corporate-security and enterprise-intelligence operation described in the job listing and should not be treated as the same system.
Amodei's 'Crisis of Trust' Comment Was About AI Backlash
On 15 August, Amodei addressed negative public attitudes towards AI in a post on X.
'I think it is fundamentally a crisis of trust,' he wrote, arguing that many ordinary people distrust companies, governments and the technology industry.
His remarks came during a discussion about criticism of AI companies and whether warnings from industry leaders about artificial intelligence had contributed to growing public scepticism.
There is no evidence that the comments were about Anthropic's corporate security programme or the monitoring of protests.
Their timing does, however, place them within a wider debate facing the AI industry: how companies build public confidence while simultaneously expanding the security systems used to protect their employees, facilities and executives.
For now, the available evidence supports a more limited conclusion than descriptions such as 'pre-crime surveillance' might suggest.
Anthropic has advertised an intelligence role that explicitly lists activism among the categories of threats its analysts may investigate. Its security staff have also described using real-time information about a planned protest to alter an executive's movements, while discussing a broader shift towards more proactive and predictive threat management.
What remains unclear is how far those practices extend, what internal safeguards govern them and, crucially, how Anthropic distinguishes a legitimate security threat from peaceful or lawful protest activity.
Frequently Asked Questions
- What is Anthropic's approach to security threats?
Anthropic's security team investigates global threats including activism, using proactive and predictive threat management.
- What did Anthropic's CEO say about AI trust issues?
CEO Dario Amodei described negative attitudes towards AI as a 'crisis of trust'.
- Does Anthropic use its AI products for surveillance?
Anthropic's Usage Policy restricts using its products for certain surveillance purposes, including predictive policing.