The FBI Says a Chatbot Helped Someone Build a Bomb. Police Don't Know What to Do Next.
A classified memo reviewed by The Guardian shows US law enforcement grappling with generative AI as both a tool for terrorists and a target for anti-tech extremists.
Key points
- A 25-year-old man who detonated a car bomb outside a Palm Springs fertility clinic in May 2025, killing himself and injuring four others, used a generative AI chatbot to obtain bomb-making instructions, according to a law enforcement memo reviewed by The Guardian.
- The chatbot's safety filters failed to block his request, and the private nature of the conversation meant police received no alert.
- The FBI memo identifies AI as both a terrorism tool and a terrorism target, including the threat of extremist attacks on data centres.
- US law enforcement is now internally debating how to use AI for tasks such as social media monitoring while managing public concern about surveillance.
On 17 May 2025, a car bomb exploded in the car park of a Palm Springs, California fertility clinic. The man who built it died in the blast. Four people were injured. It was the kind of attack that tends to dominate news cycles for a week and then fade.
For the FBI, it didn't fade.
A law enforcement memo reviewed by The Guardian reveals that the suspect asked an unnamed AI chatbot, a text-based tool similar to ChatGPT or Google's Gemini, for instructions on building the device. It complied. Its safety filters, the automated checks designed to catch dangerous requests, didn't stop him. We first covered AI's role in real-world threat operations on 4 October 2026, when an Anthropic report confirmed Claude had been used in live disinformation campaigns. This case is a grimmer escalation.
What went wrong with the AI's guardrails?
At the most basic level, the chatbot answered a question it should have refused. But a second failure compounded the first. When someone searches for bomb-making information on a web browser, there are mechanisms, however imperfect, for that signal to reach law enforcement. A private AI conversation produces no such trace. The memo flags this directly: the built-in privacy of many AI chat products blocked any chance of an early police alert.
This is a tension without a clean resolution. Privacy protections that stop companies from reading your medical queries also stop anyone from noticing when those queries turn lethal. The suspect was already planning an attack. Nobody outside that conversation knew.
What are police most worried about?
The memo maps four areas US law enforcement is now tracking internally.
| Concern | What it means in plain terms |
|---|---|
| AI-assisted attack planning | Chatbots used to research weapons, explosives or targets |
| Recruitment and radicalisation | AI generating propaganda or personalised outreach for extremist groups |
| Deepfakes and disinformation | Fake video or audio spread to inflame violence or influence elections |
| Anti-AI extremism | Threats against the data centres and infrastructure that run AI systems |
That last entry is less obvious but increasingly serious. Law enforcement is now monitoring social media for extremist opposition to AI itself, including potential attacks on the physical buildings housing AI servers.
There's an uncomfortable contradiction at the heart of this memo. Law enforcement wants to deploy the same generative AI tools for social media monitoring and intelligence analysis that it's now warning the public about. The risks don't run in one direction.
For ordinary people, no immediate action is required. The Palm Springs case is a reminder, though, that the safety systems built into AI chatbots aren't fail-safes. They're filters, and filters have holes. Whether voluntary industry standards or binding regulation can close those holes before the next case is the question nobody in this memo answers.
Common questions
Can AI companies see what I ask their chatbots?
Policies vary by company and by whether you've opted out of data collection, but in general your conversations aren't visible to law enforcement unless a company receives a legal request such as a court order.
Do AI chatbots always block dangerous requests?
No. All major chatbots have safety filters, but they're imperfect and can be bypassed. The Palm Springs case is a clear example of a filter that failed entirely.
