It seems like just recently I was writing about the so-called 10% chance of the destruction of humanity due to the race between AI companies, and now there is another scandal around these giants, specifically Anthropic. Namely, the development of software for missiles, as well as the creation of their own biology lab. By the way, the publication of the report that will be discussed in this article directly coincides with the story of the researcher Jacob Coxon, whom I wrote about recently.  

The 154-Page Report and Biological Threats  

So, on September 11, Anthropic published a large 154-page report documenting the misuse of its AI models, such as Claude Haiku, Sonnet, and Opus, for the period from December 2025 to August 2026.  

There were 5 specific cases recorded where scientists or bad actors tried to use Claude for tasks directly related to biological weapons. Namely, attempts to use AI to modify the Chikungunya virus and avian influenza in order to increase their danger. Anthropic itself evaluated the so-called uplift (this is a kind of increase in capability)—shortly speaking, it is how much the AI eases the task for a bad actor compared to a regular internet search. Ultimately, they came to the conclusion that advanced models (advanced artificial intelligence models), without proper security, will make creating pathogens much easier and cheaper.

Drone Warfare, Cyber Espionage, and Hackers  

Moreover, as I mentioned at the beginning, besides the biological lab, the report mentions that there were attempts to connect Claude to the development of software for targeting missile systems and drones. For example, attempts were recorded from the territory of Northern Yemen, which is most likely linked to the Houthis. It was also used for cyber espionage, disinformation campaigns, and propaganda, which is already more connected to Russia, China, Turkey, and Iran.  

There is even already an example where the Russian hacker group CyberAsylum used AI to fully automate phishing campaigns. Attacks were carried out on more than 20 organizations in Ukraine and Europe. Separately, a group of Russian freelance developers used Claude to create software for the autonomous control of a swarm of kamikaze drones with target recognition.  

Anthropic’s Secret Biology Lab  

And it seems like after so many not-very-good use cases from a company that deals with AI development, a week after the report, Reuters and CNBC confirmed that Anthropic secretly built and launched its own physical lab in the San Francisco area. The head of biological sciences at Anthropic, Eric Kauderer-Abrams, confirmed that digital modeling is no longer enough, quoting his words: “When you want to do biology, the ultimate test remains actual lab work, and that will be true for a long time.” The lab is already working with real biological samples and DNA.

The ultimate goal of Anthropic is to train Claude to manage the lab without human help. It all works in the following way:

• AI thinks - invents an experiment.
• Robot does - automatically performs experiments in a real laboratory.
• AI analyzes - evaluates the result and immediately plans the next step.

P.s. I hope I explained this principle well.  

Big Pharma Contracts and Strategy  

Also, Anthropic previously secretly bought the startup Coefficient Bio for about $400 million in stock. The company also added Vas Narasimhan, the CEO of the large pharmaceutical giant Novartis, to its board of directors.  

Although Anthropic already has commercial AI contracts with giant pharma companies, for example, Novo Nordisk, Bristol-Myers Squibb, and Genentech/Roche, Kauderer-Abrams emphasized that they do not plan to enter the clinical trial stage and compete with clients. Instead, Anthropic will focus on the early preclinical phases of drug development for rare or so-called "undruggable" diseases, which are economically unprofitable for big pharma companies to deal with due to the small market.  

Dario Amodei’s Essay and Geopolitics  

And by the way, the CEO of Anthropic, Dario Amodei, recently published an essay where he stated that AI and the automation of biology will allow humanity to cure most major diseases within the next 5–10 years, shrinking medical progress quite a bit. Anthropic’s management believes the threat is not the artificial intelligence technology in biology itself, but the uncontrolled access of the general public to dangerous knowledge, like step-by-step instructions for synthesizing viruses in a regular chat. They believe that such work and experiments should be conducted in closed, licensed, physical laboratories under the strict supervision of experts. Which I basically completely agree with.  

Also, Amodei openly calls on the US government to speed up the certification of drugs created with the help of AI, but at the same time to tighten export controls on advanced chips.  

The China Factor and the Final Verdict  

His main argument is that the US cannot afford to slow down AI research because of the fear of risks, since geopolitical rivals Russia and China (although I think China will pose a bigger threat to them here) will move forward without any internal security restrictions. And since I started talking about China, I want to add to the text that Chinese state institutes used bypass servers in Singapore to test their own AI models through Claude Sonnet I think most of you know models like DeepSeek or Qwen. That is, China is using American AI as a "trainer" for its models.  

And to conclude, I will add not in my own words, but with a thesis from Silicon Valley, which in my opinion best describes the race in this contest: the US can control its own companies, but it is unable to stop world progress. If the Americans hit the brakes out of fear of risks, China will definitely not stop with them. I will only add from myself that the world of AI is changing way too fast.  

Thank you everyone for reading this material.