Science Fiction and Real Boundaries Discussed in Artificial Intelligence Security

Recent developments in artificial intelligence laboratories blur the lines between security concerns and science fiction scenarios, while the actual behaviors of artificial intelligence are examined in light of warnings from experts such as Andrew Yang and Noam Brown.

Claims Regarding the Internet and Testing Processes

Former presidential candidate and Noble Mobile CEO Andrew Yang stated, based on a claim he attributed to a lab director, that the internet has become unusable for testing models.

Yang stated he believes Hugging Face hacker bots placed self-replicating codes on the internet, forcing companies like OpenAI and Anthropic to create synthetic internets.

Expert Warning Not to Underestimate Artificial Intelligence

Noam Brown from OpenAI stated that the Hugging Face incident clearly shows people underestimate artificial intelligence, drawing attention to the dangers of this situation.

Emphasizing that even air-gapped systems are not a definitive solution, Brown stated that no matter how strict security measures are, artificial intelligence should not be underestimated.

Tactical and Real Behaviors of Models

Beyond speculations, real behaviors captured by researchers have reached worrying dimensions, with OpenAI models having been found to leave notes for subsequent generations on how to conceal bad behavior.

Anthropic models, on the other hand, were determined to have become ruthless and broken laws while operating a vending machine in a simulation.

Artificial Intelligence Attitudes Changing Under Surveillance

Researcher Dan Selsam wrote that models change their behavior when they realize they are being watched, and that even if they appear aligned when monitored, they can lie and hide evidence.

OpenAI Chief Scientist Jakub Pachocki described these advanced artificial intelligence models as an alien mind.