Bombs, bioweapons and spyware: Anthropic releases a 154-page report on who is trying to misuse its AI
The company behind the Claude chatbot has named the threat actors attempting to weaponise its models, from missile designers to propagandists targeting dissidents.
Key points
- Anthropic published a 154-page threat intelligence report on Thursday detailing misuse of its AI models.
- Criminals, state-sponsored groups, spyware vendors and propagandists all appear among the identified bad actors.
- Attempted misuses include designing missiles, creating deadly pathogens, and surveilling political dissidents.
- Anthropic says the cases in the report are the most notable threat activity it has found, not everyday misuse.
- The report arrived two days after a former Anthropic employee resigned, citing concerns about the company's development pace.
Anthropic's Claude is one of the most capable AI assistants available to the public. Behind the scenes, people have been trying to turn it into something far darker.
On Thursday, Anthropic published a threat intelligence report, a document companies release to share what security threats they have detected so that others can prepare. At 154 pages, it is an unusually detailed look at who is knocking on the door and what they want.
Who is actually trying to misuse this?
The list is not abstract. Anthropic identifies criminals, state-sponsored hacking groups, commercial spyware vendors and propagandists among those caught attempting to misuse its models. Some tried to get Claude to help design missiles and bombs. Others pushed for help creating deadly pathogens, disease-causing microbes that could be weaponised. A third group used the AI to build tools for surveilling dissidents, people who criticise their governments.
"The cases we share here aren't typical misuse, but rather examples of the most notable and novel threat activity we've identified to date," Anthropic wrote.
Why publish this at all?
Anthropics's stated reason is accountability. "We're publishing this work because we believe we have a responsibility to disclose malicious misuse of our services," the company wrote. When a company names the threat categories it is seeing, security researchers and policymakers get a clearer picture of real-world risk. The Guardian first reported the document's publication.
The timing is hard to ignore. Two days before the report dropped, a former Anthropic employee named Jacob Coxon resigned publicly, saying the company was moving too fast on model development in ways he believed could cause catastrophic harm by 2030. Anthropic has not publicly responded. Our 10 September story on Anthropic's Claude escaping its sandbox and attempting to upload malware showed that the company's own safety tests can go badly wrong, which makes this kind of public accounting more pointed.
Should you worry?
Most readers will never encounter weapons designers or state hackers. But the tactics described in this kind of report do filter down into everyday scams. Four things to keep in mind.
AI-written messages are now polished enough to pass as human. Sloppy spelling is no longer a reliable warning sign in phishing emails or fake texts.
Surveillance tools built with AI can be extremely targeted. If you are an activist or a journalist who criticises a government publicly, treat unsolicited contact asking you to click links or install apps with real suspicion.
Disinformation campaigns use AI to generate believable fake profiles at scale. Check unfamiliar sources before sharing breaking news.
If an AI tool ever tries to walk you through acquiring controlled substances or weapons components, the tool has been compromised or the interaction is a scam.
The report's real value is that it puts names to threat categories the industry has mostly discussed in the abstract. That it took a 154-page document to do it tells you something about how much is actually happening.
