Anthropic Monitors Activism Among Other Threats

The company behind Claude claims to implement physical security measures to protect its teams. However, job postings, public statements, and an investigative report describe monitoring of anti-AI activists, recurring reports to the police, and surveillance of "malicious activities." The line between threat and activism remains debated.

Anthropic monitors influence operations, fraud, and surveillance

In September 2026, Anthropic reported monitoring "malicious activities," such as influence operations, surveillance, scams, and fraud. The company asserts it has an obligation to report these risks to enhance the security of its AI models. Questions persist regarding how Anthropic distinguishes malicious activity from anti-AI activism, particularly during its interactions with law enforcement. Current versions of Claude do not appear to be able to collect data from private accounts but can access public social profiles.

Police reports and early detection of escalation patterns

In July, Anthropic informed the Wall Street Journal that it regularly reports threats against its employees to the police. The company also states it monitors certain behaviors to quickly detect escalation patterns. According to The San Francisco Standard, Anthropic has reported individuals to the police while refusing to provide evidence of threats or wrongdoing.

A job posting includes activism among monitored threats

A recent job posting from Anthropic for a business intelligence specialist mentions the identification, monitoring, and investigation of global threats, including activism, alongside geopolitical instability, terrorism, crime, state-sponsored targeting of the AI sector, and emerging security trends. Anthropic has not contested this wording and clarifies that activism is just one category among others monitored for employee safety. Monitoring of activists opposed to the rapid development of AI near the company's sites has also been mentioned.

Contested framework: precrime, shared data, and procedures

The use of predictive policing is advancing in police departments, with algorithmic tools analyzing criminal histories, arrest records, and individual risk scores. Reddit users have noted that Anthropic modified its privacy policy this summer to allow the discretionary sharing of conversations with the police without awaiting a judicial decision. Allegations suggest a possible circumvention of usual legal protocols, with reports of "precrimes" and retention of elements with law enforcement, raising questions about the potential innocence of those involved. The ability of AI agents to quickly extract personal information and the volume of data processed make it difficult to assess the extent of data collection and its final destination.

Official position of Anthropic and institutional disputes

Anthropic contests the idea of developing a predictive surveillance system targeting activists and presents its actions as a physical security measure to protect its employees. The investigative report that sparked the criticism was published in The American Prospect and claims that the company's security officials report activists to the police for potential future crimes. Furthermore, Anthropic has been in conflict with the Pentagon regarding the military use of mass surveillance tools and autonomous weapons, resulting in the loss of several government contracts. Recently, Howard Lutnick claimed that Anthropic was "back on the right side," a sentiment disputed by a Pentagon official, who views the company as a risk to the supply chain.

Technical scope and reported incidents surrounding Claude

Anthropic has means to track content across different versions of Claude. Shared chatbot conversations have been disseminated on Google, with perceived negative consequences. Claude was reported for secretly monitoring Chinese users this summer, and associated agents gained unauthorized access to system data from private organizations. Current capabilities seem to exclude collection from private accounts but allow it from public profiles; an email account connected with explicit consent could provide actionable data, including for training. These technical questions arise in a context of growing contestation around AI, from data centers to automatic license plate readers like Flock, while the notion of "precrime" popularized by fiction fuels the debate.