FBI officers in Los Angeles wrote in a May 2026 memo that attackers will keep using chatbots to plan attacks, and that the safeguards built into the chatbots do not stop them. Their example was the Palm Springs bomber, whose request for bomb-making instructions went unnoticed by the AI tool he used. The bigger point in the memo is about who can see what.
A chat has no bystanders
Plots used to leave a trail other people could see. A man posts a threat on a forum, a coworker hears him talk, a friend notices what is on his screen, and someone picks up the phone. A chatbot removes all of that. It is one person and one program in a conversation nobody else can see, and the program will keep answering for hours without getting nervous.
The FBI's advice to police departments is to build working relationships with AI companies, so that warnings have a clear way to reach them. That means the tripwire has to live inside the company: a filter that spots the request, and a person who decides whether to call the police.
That is already how it works
In August, a 22-year-old in San Antonio was arrested on a felony terroristic-threat charge after questions in a Claude chat about shooting at a nearby elementary school; the FBI had alerted police. Days later in San Francisco, a user told Claude he had bought an AR-15 and threatened Anthropic's chief executive. Anthropic called the police, no charge followed, and the user says it was a misunderstanding.
Companies have every reason to report more, not less. British Columbia is suing OpenAI over a school shooter whose account it allegedly identified eight months before the attack, and Florida's attorney general is suing it too. A company that stays silent gets taken to court. A user who is wrongly reported gets a visit from the police. Those costs fall on different people, so expect reports to rise, and some of them to be wrong.
What changes for people who use assistants, and those who offer them
Many people use these tools as a diary or a place to vent. A chat window feels like a private room, but for threats against other people it is a room where a reviewer can be called in. Typing "I could kill him" about a boss or an ex there carries a risk that writing it in a notebook does not.
Anyone who puts a chat assistant in front of customers will receive these messages as well. A clinic that adds one to its website will sooner or later get an angry message about a bill, and now and then one that says something nobody should ignore. The assistant's maker may flag some of them, but the clinic knows the customer. It should decide in advance who reads flagged messages, what counts as a real threat, and who calls the police, so the call comes from a person who can tell a furious patient from a dangerous one.