Anthropic Called SFPD, Then Withheld the Claude Chat

Anthropic spent the last year selling itself as the lab that would not build mass domestic surveillance for the Pentagon. This month its own security shop is cataloging “activism” beside terrorism on an intelligence brief, and its chatbot team is calling the San Francisco police about Claude users while withholding the transcript.

Daniel Boguslaw at The American Prospect reported the architecture on September 9 from job postings and interviews. Anthropic contracts with Samdesk to watch protests near executives and offices. Global Security Operations Center manager Keon Ellison described a one-hour heads-up that let staff reroute an executive through a hotel service door. Security program language is blunter. The goal, as Prospect quotes it, is “proactive and predictive and preventative threat engagement.” A Global Safety, Intelligence, and Security posting last month listed activism among the global threats an $180,000-to-$230,000 specialist would “identify, assess, track, and investigate,” with OSINT on specific actors.

6 min read

The Wall Street Journal heard the same posture in July: a person-of-interest process to “catch escalation patterns early.” Several people later involved in police incidents, the Journal said, were already in Anthropic’s file.

The chat the cops did not get to read

The chat-to-cops pipeline showed its working parts on August 14. A man told Claude he had bought an AR-15 and had Dario Amodei “in his sights,” according to a police report obtained by The San Francisco Standard. Anthropic waited until the following Tuesday, then called SFPD, which sent officers to 500 Howard Street and found nobody. The employee who filed the report, the officer wrote, “refused to show me the messages due to Anthropic’s company policy.” The man later told the Standard he was “just fucking around.” He has not been arrested or charged. An Anthropic spokesperson called the sequence “our safeguards process working as intended.”

Real violence against this industry is not a rumor. Someone threw a Molotov at Sam Altman’s house in April. Guards stopped a man in Anthropic’s lobby days later. Liferaft told the Standard that threats against AI leaders and data centers rose sevenfold from February to May. Eric Goldman at Santa Clara put the legal squeeze cleanly: companies fear liability if they stay quiet, and they over-report as they get more hated.

The problem is the category jump. A rifle-and-CEO chat is a threat review. Listing activism on the same intelligence menu as terrorism is a political file. Predictive protest tracking is how a lab that won a First Amendment fight against a Pentagon blacklist starts acting like a miniature fusion center. Amodei wrote last month that ordinary people “always suspect that we are cooking up some new way to screw them over.” He diagnosed a crisis of trust. Then his security team asked for earlier notice of the people who distrust him in public.

Last week the guards who walk OpenAI and Anthropic campuses authorized a strike over $22 an hour. Anthropic told staff to work from home. The people paid to stand in the doorway are cheaper than the people paid to predict who might show up there.