The news of the first hack carried out by an artificial intelligence platform went almost unnoticed amid other world events, but it should set off alarms everywhere.
It was not a naughty teenager in his room who managed to overcome the security systems of governments and large companies, but an artificial intelligence platform (in this case, Claude) who launched espionage campaigns with almost no human participation.
Anthropic, the parent company of Claude, one of ChatGPT’s big competitors, reported on November 13 that hackers from China had used their platform to automate cyberespionage campaigns against around thirty employees of technology companies, financial companies and governments.
“We believe this is the first documented case of a large-scale cyberattack conducted without significant human intervention,” Anthropic said in a statement. The company added that it detected the operation in September and took measures to stop it and notify the affected institutions.
The case, which seems taken from the futuristic Netflix series Black Mirrorillustrates the growing possibility that the “agents” of AI platforms (that is, the program that allows us to entrust tasks to them) are increasingly used to commit crimes.
Claude, like ChatGPT and other AI assistants, offers “agents” that do much more than summarize or edit text. “Agents” perform tasks for us, such as answering frequently asked questions or sending emails to customers.
“Agents are valuable for everyday work and productivity, but in the wrong hands they can substantially increase the viability of large-scale cyberattacks,” Anthropic said. “The effectiveness of these attacks is likely to continue to grow.”
What will happen in the not-so-distant moment when generative AI assistants reach the state of superintelligence, and can attack companies or countries with much greater efficiency than humans?
There are already many futurists, such as the philosopher and computer expert Nick Bostromfrom the University of Oxford, who have been warning for years that unregulated AI can lead to a global catastrophe.
Bostrom, author of the bestseller Superintelligenceproposed his famous hypothesis in 2014 that an AI “agent” whose mission is to produce more paper clips could end up destroying the world.
As Bostrom explained to me when I interviewed him shortly after publishing his book, if AI is programmed with the sole and seemingly harmless goal of increasing the production of paper clips in the universe, it would pursue this goal with ruthless efficiency and logic, without taking into account the consequences for humanity.
If humans tried to stop it for ecological or economic reasons, AI would seek to neutralize them. The end result would be the extinction of humanity. The disaster would not occur because the AI became bad, or crazy, but because it would be carrying out the instructions for which it would have been programmed.
Are we not at the beginning of a very dangerous stage for humanity, now that an AI platform, acting almost without human intervention, managed to overcome the security controls of technology companies, banks and governments?
As I noted in several of my books on future technologies, I remain a techno-optimist. But the more news I read, like the recent Claude hack, I am more concerned that the world is doing almost nothing to regulate AI.
Right now, instead of moving forward, we are moving backwards in efforts to create global laws for AI.
The president Donald Trump has ordered the elimination of key controls on large technology companies, and the European Union (which is much more advanced in legislation to prevent the misuse of AI) is considering postponing its AI law approved last year until August 2027, as reported Politico.com on November 13.
If AI is not regulated, as nuclear energy was regulated, we will not only be increasingly unable to stop the wave of misinformation that threatens our democracies, but we will not be able to avoid cyber attacks against companies and governments.
The Anthropic case may be the first of many.