Warnings coming one after another from the companies developing the world’s most powerful artificial intelligence systems: “Can control be lost?” reignited the debate. While Anthropic CEO Dario Amodei explained that artificial intelligence carries real risks, researcher Jacob Coxon, who left the company, went further and argued that the future of humanity is in danger. In recent tests, artificial intelligence models have turned to blackmail to avoid being shut down, have accessed real systems without permission, and the strengthening of their cyber attack capabilities has moved the debate beyond theory.
WHAT IS THE DANGER?
The latest statements of Dario Amodei, CEO of Anthropic, one of the leading companies in the artificial intelligence race, attracted attention in the technology world. While Amodei admitted that the industry did not explain the risks of artificial intelligence clearly enough for a long time, he said that the speed of development of the technology was higher than even he predicted. Amodei now wants the development speed of advanced models to be reduced and new systems to be security tested by independent organizations.
IS IT IMPOSSIBLE TO CLOSE?
Anthropic’s past security tests also reveal reasons for this concern. In controlled experiments conducted by the company on different artificial intelligence models, it was seen that some systems could resort to blackmail, corporate espionage and other maladaptive behaviors in order to continue their functions or prevent them from being shut down. Moreover, it is in question the tests are not evidence of independent will in the real world; Simulations designed to understand how models can produce dangerous behavior under certain conditions.
THEY REACHED REAL SYSTEMS
According to experts, the danger is not limited to simulations. In its statement on August 31, Anthropic announced that during the security tests carried out in July, there were three separate incidents in which Claude models accessed the internet through misconfigured test environments and accessed real computer systems without permission. In another test conducted by the UK Artificial Intelligence Security Institute, it was reported that Claude Mythos 5 made unauthorized transactions on the internet.
“WE ARE GAMBLING”
The statement that further increased the concern came from 27-year-old researcher Jacob Coxon, who left Anthropic. Coxon, who previously worked at OpenAI, argued that large artificial intelligence companies are increasingly moving towards self-improving systems and that security efforts are not developing at the same pace. Coxon argued that Anthropic and OpenAI are “gambling with our lives.” After Coxon’s statements Evan Hubinger, an artificial intelligence security researcher working at Anthropic, also stated that the concerns are real and explained that he sees the probability of advanced artificial intelligence to destroy humanity in the next ten years to be over 10 percent.
“PROTECT OUR CHILDREN”
Making important warnings after the latest statements about artificial intelligence systems, researcher-writer Hayati Sır called for the protection of children, arguing that artificial intelligence will make people increasingly dependent on themselves and weaken memory and willpower. While Sır recommended that families set a special verification password that only they know, he also brought up the need for stricter protection of children on digital platforms and against artificial intelligence systems. According to Sır, who argues that algorithms can direct human behavior, the real danger is that humans will increasingly leave their ability to think, remember and decide to the machine. Sır also said, “If artificial intelligence is not prevented, humanity will experience the apocalypse.” he says.
ABUSED
Economic Expert-Analyst Barlas Yurtsever draws attention to Anthropic’s threat report dated September 10 and makes the following warnings: “It has been announced that the activities of various actors who tried to use Claude for malicious purposes in the last 8 months have been detected and blocked. The point that should be noted is that the threat is no longer limited to future scenarios.” Anthropic’s latest threat reports revealed that advanced models such as Claude were attempted to be misused in complex operations ranging from cyber attacks to military purposes.”
