Anthropic reports to police a threat against Amodei detected in Claude: chat privacy has limits

By Zendoric · September 5, 2026.

On August 14, a user wrote in a conversation with Claude —Anthropic's artificial intelligence assistant— that he had bought an AR-15 semi-automatic rifle and that he had the company's chief executive, Dario Amodei, "in his sights", according to a police report obtained by The San Francisco Standard. Four days later, the following Tuesday, Anthropic alerted the San Francisco Police Department (SFPD), which sent officers to the company's headquarters at 500 Howard Street and did not find the person in question there.

Reached by phone by the Standard, the man —who has not been arrested or charged with any crime— said he was "just joking" and appeared embarrassed by the episode. Anthropic confirmed that it "banned the account, in line with its standard practice, and referred the case to the authorities", and described what happened as its "safeguards process working as intended".

The employee who passed the alert to the police said he did not feel in imminent danger, but wanted the incident on the official record. According to the report, he refused to show officers the exact content of the conversation "because of company policy" — a detail that captures the underlying tension: Anthropic acts as the complainant, but keeps under lock and key the very material that prompted the complaint.

The episode is not an isolated case. In April, the Russian Hill home of OpenAI chief executive Sam Altman was targeted with a Molotov cocktail. Days later, as The Wall Street Journal reported, a man was intercepted in the lobby of Anthropic's offices after claiming that a company executive would be killed. Threats against AI executives and data centers increased sevenfold between February and May, according to threat intelligence firm Liferaft. Palantir, Oracle and Salesforce have increased their spending on executive protection over the past year, the Standard reports.

The case also exposes a legal and privacy gray area. Eric Goldman, a law professor at Santa Clara University, told the Standard that internet companies "face significant liability if they fail to report and a crime is in fact committed", but warned that, as these companies become more unpopular, "there is more pressure to over-report, knowing that some of the people flagged should not be flagged at all". Claude's privacy policy allows Anthropic to report to the police any threat of violence detected in a conversation; the company said its automated systems detect and block this type of exchange, but declined to clarify whether it treats threats against its own employees differently from other mentions of weapons or attacks.

Anthropic is also dealing with another source of tension at its San Francisco headquarters: it has asked employees to work from home in anticipation of a possible strike by the building's security guards, subcontracted through Allied Universal, whose collective agreement expired in June without a renewal deal.

Our reading: this particular episode will probably be resolved as what everything suggests it is —a tasteless joke that got out of hand— but the mechanism it has set in motion deserves more attention than the case itself. AI chatbots have become, almost without any prior public debate, a channel capable of referring conversations that users assumed were private directly to the police. It is a break with how we have understood digital privacy until now: searching on Google or posting on social media always carried an assumed public or semi-public component, but talking to a conversational assistant is perceived —wrongly, as we have just seen— as a closed space.

The expert quoted by the Standard sums it up well: the more widespread social rejection of the big AI companies becomes, the more incentive they have to over-report rather than under-report, even if that means flagging innocent people. It is a logic of defensive legal liability, not of well-calibrated public safety, and the result looks more like "Minority Report" —the comparison the article itself makes— than a system with safeguards: action is taken on a sentence, not on an accomplished fact.

This connects with the thesis we have been defending at Zendoric about the transition to AI: the short term is hard, and not only because of jobs. The social backlash against these companies —over the impact on employment, the use of creative works without consent, the energy footprint— feeds on episodes of real violence against their executives, which in turn pushes the companies to monitor the use of their own products more closely. It is a circle that erodes trust in both directions: users trust the privacy of their chats less, and companies trust their own users less.

In the long run we remain convinced that the abundance AI can bring —in health, in time, in productivity— will more than offset these costs of the transition. But that future requires the companies building it to survive politically in the present, and that calls for transparent, proportionate moderation protocols, not the opaque discretion that today no one —not even regulators— can audit. Anthropic says its "safeguards process" worked as intended; the problem is that, outside the company itself, nobody really knows what that process consists of.