The latest September Anthropic study on the misuse of their AI models is simply a real mine of past that Tom Clancy himself would not be ashamed of, the difference is that it truly happened.
A race of armaments in artificial intelligence can be brutal, but seldom so hilarious. Chinese technology companies, wishing to catch up with American models at a inexpensive cost, quietly redirected queries to Anthropic servers. Effect? Nothing ignorant (?) military engineers, police officers and employees of the Chinese and Russian government voluntarily sent e.g. secret surveillance footage and active passwords to state databases.
Chinese startup steals, Chinese army pays for mistakes
Moonshot AI, the creator of Kimi, alternatively of processing client questions at home, sent them straight to Claude. It was about alleged illegal distillation – stealing the way the American algorithm was reasoned to train its own product on it. Kimi users, however, were absolutely convinced that they were utilizing safe (in their opinion) in a 100 percent of Chinese software. They didn't know anything about the redirect.
In this way, a user associated with the Chinese Army (PLA) asked artificial intelligence to analyse CCTV surveillance recordings in Chengdu for different behaviors of the tracked person. The strategy received data from hundreds of cameras, including those placed in front of military facilities and state institutes. In turn, the engineer of a large state-owned company pasted the interior code and active log-in data of many key Chinese technology companies into the hut.
Passwords for the Russian military on an American server
Identical tactics were applied by another giant, DeepSeek. Redirecting queries to Claude to collect training data, the company unwittingly exposed the secrets of its own citizens and geopolitical allies. Caught in this computer trap, working with data from a Russian government agency affiliated with the Ministry of Defence of the Russian Federation, he pasted into a supposedly Chinese chat room... active passwords to the Russian government database. With the eyes of my imagination, I can see in Langley (CIA) and Fort Meade (NSA) beginning champagne.
There's no end to the depictive disasters. The employees of the Chinese technology company utilized the hut to analyse the full specifications, organisational structure and strategical objectives of the flagship artificial intelligence program. Worse still, Chinese engineers building a case management strategy for the local Public safety Office utilized DeepSeek to make a tool to track citizens after ID numbers. The full surveillance architecture went consecutive to the logs of the American corporation.
The Anthropic study exposes the gigantic paradox of the modern technological race. The authoritarian regimes pump billions into the construction of “syphernical” artificial intelligence, while their own corporations—out of pure pragmatism and profit desire—are cut short, stealing Western technologies to get better results (the mentioned distillation training based on the results of a competitive AI model). In effect, alternatively of protecting national secrets, they created the world's most perfect free Intelligence vacuum cleaner working consecutive into the American cloud. If you want to read the study yourself, I'll mention you to Anthropic resource.
Cloud AI is simply a mystery vacuum cleaner. Samsung escapes from Silicon Valley and locks Mistrala in the basement
If article They stole AI to catch up with the West. They accidentally gave state and military secrets to Americans does not look right in your RSS reader, then see it on iMagazine.
