Artificial intelligence, which was supposed to help write code and analyze data, has become a tool that intelligence services are fighting over. Anthropic, the developer of the Claude model, confirmed: malicious actors and state-affiliated groups in China and Russia have already attempted to use its AI for cyberattacks, mass surveillance, and research bordering on the creation of biological weapons.
This is not a hypothetical science fiction scenario, but a documented fact in the company's new report, as Politico writes. And this is where an unexpected twist emerges: protection from AI abuse now lies not with states or international organizations, but with a private company from Silicon Valley, which itself decides what to block and what to allow.
When a company's filter becomes the front line
Anthropic has documented attempts to automate cyberattacks and intelligence operations, some of which were linked to state structures in China and Russia. This means that conventional cybersecurity methods — firewalls, antivirus software, network monitoring — are supplemented by a new layer: moderation of queries to the AI itself.
The practical conclusion is simple: companies developing powerful models are becoming de facto counterintelligence agencies. They see requests before those requests turn into actual attacks, and have the resources to block them — or not block them, if they fail to recognize the threat in time.
The company said it blocked an attempt by a user to get help with research that could have made a dangerous virus more harmful.
This episode with a biological threat shows where technology stops being simply a productivity tool. If a model is capable of helping make a virus more dangerous, it is theoretically capable of helping in other areas where years of specialized training and entire laboratories of specialists were previously required.
Copying as a parallel threat
In addition to direct attempts at abuse, Anthropic is fighting a large-scale illegal ecosystem whose participants attempt to bypass the company's restrictions by creating numerous fake accounts to access the technology. The goal is to illegally copy Claude's capabilities and use them to train other, less controlled models.
This creates a practical problem for any user or business: the stricter the restrictions a company imposes in response to abuse, the more difficult and slower access to the service becomes for ordinary customers. Protection from state hackers inevitably affects speed and convenience for everyone else.
What this means for the balance of power
The company directly warns: as AI capabilities grow, such systems will increasingly simplify conducting complex cyberattacks that previously required entire teams of specialists. This changes the very economics of cyber threats — the barrier to entry for conducting a serious attack is falling, not rising.
It is telling that these accusations came against the backdrop of mutual claims between Washington and Beijing: China rejected U.S. accusations of harmful AI development ahead of a Trump-Xi meeting, and later threatened retaliation over accusations of technology theft, although it did offer negotiations. The AI race is becoming yet another front in the confrontation between major powers, where private companies find themselves between the lines of fire.
The question that remains open: if private AI developers are already performing the function of a national security filter, who will oversee their own decisions about what constitutes a threat — and will regulators be able to keep pace with the speed at which these technologies are spreading?
