Anthropic, the company behind the AI Claude, recently revealed a catalog of real cases where its own technology was used for some fairly disturbing purposes. And by disturbing, I mean espionage, mass surveillance, fraud, and even weapons development.
What makes this kind of disclosure different is that it doesn't come from an outside researcher or a leak: it's the company itself admitting, with code names and operational details, that its technology was used against its own policies. And that, more than being reassuring, made me think about how fast this landscape is shifting.
When AI stops being an assistant and becomes an operator
For a long time, the idea of "using AI to cause harm" was imagined as something like using a smarter search engine: asking for advice, shortcuts, the occasional tip. But what's described here is different. A group was identified, with indications of links to Russian espionage, that fully automated the lifecycle of its malware: whenever a security tool detected its software, an AI-based system would rebuild and redeploy it without direct human intervention.
That changes the nature of the problem. It's no longer someone leaning on a tool, but an operation that sustains and repairs itself.
Manipulating public conversation
Another section documents influence operations: networks of fake social media profiles designed to simulate consensus where none exists. Hundreds of accounts "voicing" the same opinion at the same time, with nothing on the surface revealing that a single coordinated operation is behind it.
What weighs most is that several of these campaigns were timed to coincide with electoral processes in different countries. This isn't random noise: it's a strategy with deliberate timing.
Surveillance at a scale that once required entire institutions
This is, for me, the hardest section to read. It documents governments and private contractors using AI to build systems capable of intercepting communications at a national level, or to generate risk profiles on dissidents. Tasks that once required entire teams of analysts can now be sustained by a single person with the help of an AI model.
This isn't a projection into the future. It's a description of something that already happened.
Weapons development
This isn't the kind of section you'd expect in a report published by an artificial intelligence company, but it's there: cases of people using AI to develop software linked to guided rockets, drones, and electronic warfare systems. The accounts involved were banned as soon as the activity was detected, though that itself leaves an uncomfortable question about how many similar cases go undetected in time.
Biological risk: the most significant admission
Of the entire catalog, this is the part that carries the most weight. With earlier generations of models, the company could state with relative confidence that they weren't capable of meaningfully assisting in dangerous biological research. With the most recent models, that certainty no longer holds the same way, which is why restrictions on that type of query have been strengthened.
It's a rare admission: publicly acknowledging that the product's own progress narrows the safety margin that was previously taken for granted.
Fraud at an industrial scale
The final section documents an organized fraud case: a network operating dozens of dating apps, with thousands of AI-generated profiles conversing with real users, combined with human workers to give the appearance of authenticity. A deception operation designed and run with the same logic as a production line.
What remains after seeing all of this together
If there's a pattern running through all these cases, it's that AI is narrowing the gap between what a state-backed actor can do and what someone with far fewer resources can do: a laptop, access to a model, and the willingness to misuse it. Capabilities that once required budgets, personnel, and infrastructure can now be replicated with a fraction of those resources.
I don't think the point is to spark alarm, but to properly size up the moment we're in: these operations aren't hypotheticals or distant risk scenarios. They already happened, they were already detected, and it's reasonable to assume they aren't the only ones.
If you'd like to check out the document, click here: Detecting and countering misuse of AI: September 2026
Author: Edu Quijano