For thirty years, the NSA’s problem was not access. The access problem was solved over time. Cables were tapped. Providers were convinced or forced. Encryption was broken, or a way around it was found. As the NSA budget grew, the collected data grew too.
But another problem was never solved. Everything they collected was the trace of an action. Who called who. Which number connected to which cell tower. Which money went to which account. Which flight was booked. We can make this list longer. All of these are records that appear after something happens. The job of the intelligence analyst was to go backwards from these traces and try to guess the purpose.
This work did not make much sense. Most of the time it became a reactive process, and it started after the event. Also, the intention behind the data could be very complicated. A person who downloads an important chemistry document can be a terrorist, but can also be a PhD student. A person who flew the same route three times can be preparing an operation, but can also be visiting relatives. In the end, these traces do not tell you the difference. The thing that tells you the difference is why the person did it, and that was very hard to find.
The data was not really missing. But it always came late.
The NSA is the agency responsible for signals intelligence in the United States. It was founded in 1952 and for many years its existence was not officially accepted. Its job is to collect and analyse communication traffic around the world. It does not work with agents like the CIA. It works with cables, satellites and code breaking techniques.
The missing piece
Now let us think together.
As I wrote above, there is a problem, and the solution is not collecting more data. It is not better classification either. The solution would be this: the intention shows itself before the action, and if possible together with the reason. Before the plan is used. While the target is still being chosen.
So how can we do this? Nobody writes their intention openly on public digital platforms. Listening is not enough either, because when a person talks to another person, they do not explain the full reason. They defend themselves. They change their sentences for the other person. The NSA could not solve this in thirty years, and it is not clear that it can solve it in the future.
Let us stop here and look for an answer to this question. What kind of system could be a place where people write the things they have not told anyone yet, naturally and without correcting themselves?
The document
On 10 September 2026, Anthropic published its threat intelligence report for the fourth time. The document collects the findings between December 2025 and August 2026 under seven headings. Cyber operations, influence operations, surveillance, fraud, biological misuse, conventional weapon development and distillation. The document is quite long. The activities inside it are quite interesting. A Russian spying operation that targeted state institutions in Ukraine and Europe. Three named propaganda organisations from Iran. A platform in Malaysia that targeted voters by electoral district. A Russian information operation run through a radio station in the Central African Republic. A fake news line in Bangladesh for rural audiences.
If we look at it from the privacy side, it does not look like a disaster. Our first reaction is that it is good that they were caught. Right? This is probably the reaction of most of us. But the interesting part is how they were caught.
The intention stage
The answer to the question above is in the influence operations part of the report. A social media platform sees an operation only when the content is published. Anthropic says it can see the operation already at the planning stage. The reason is quite simple. While the actor plans the campaign, chooses the targets and writes the text, the actor explains all of this to the model.
Let us be careful here, because it is easy to reach the wrong conclusion from this sentence. The thing that is caught is not the meaning of a single sentence. The thing that is caught is the way the work is done. Producing content in bulk in a certain format, similar accounts, repeating flows. You do not see these by looking at one message. You see them by looking at the pattern.
Identity comes after this, through classic investigation, and it is based on three sources that match each other.
The first source is the prompts of the users themselves. Access from Iran is blocked, so they used a VPN. But inside the conversation, they told the model many times which city they were in, which institution they worked for and what their title was.
The second source is account information. More than seventy fake news sites were connected to one single source. This was not because the content was similar. It was because they shared the same distribution identity, and because the domain names were all registered together in a period of only ten weeks.
The third source is matching. Anthropic sees what the user produces, but it cannot see where the content goes after it leaves Anthropic. The report says this too. They closed this gap by searching for the produced content outside and comparing it. A post on the Sputnik Africa account matched exactly with a text produced in Claude. This third source does not give identity, but it proves that the produced content really went public.
Your capacity no longer tells us who you are
In my opinion, the strongest sentence in the report is this: the thing that separates state-backed actors from small criminal groups is no longer sophisticated moves, it is intention. We can see this very clearly in the examples too. Nobody tries to build the capacity needed for these operations. They buy it. A digital advertising agency in France and an IT company in Istanbul probably ran these operations next to their normal work. The report calls this model “influence as a service”, and there is an important detail. It also gives the customer the chance to deny everything.
You also do not need to feed an army. The operation in Bangladesh was the work of one single person. The profiling platform that targeted political institutions in Europe, together with a database of millions of rows, was also built by one single person.
When you put these side by side, the conclusion is this. You can no longer understand who is behind an operation by looking at their ability. Iran or Russia may sound normal to all of us. The real issue is that the same capacity also exists in one single person who lives in a very small city in northern Bangladesh.
The cost
Let us go back to the beginning.
The thing the NSA could not solve was not a technical problem. Knowing the intentions of people is really impossible. This was the missing piece for thirty years, and even huge budgets could not solve it.
Today that piece is in place. And there is no force, no court order, no secret program. People give their plans, their reasons and the things they have not told anyone yet to the AI as prompts, in raw form. The reason they do this is that they want a useful output. For the first time in history, the cost of collecting intelligence data is negative. In other words, the side that gives the data pays a subscription fee for it.
This mechanism works well today. The people who were caught were the people who should be caught. Also, the company does not hide this. It published it itself, with all the details. This is something intelligence agencies would never do.
But is there any guarantee that this will continue like this?
Maybe the company’s wish to keep behaving well. What happens if they do not continue?
I am waiting for your answer.