American Anthropic announced Tuesday that it cancels the launch of the new Mythos artificial intelligence model, as it proved too dangerous. The system detects and exploits security gaps in operating systems, while during the tests it managed to escape virtual environment and email the researcher. He even found 27-year-old vulnerability in OpenBSD and allowed engineers without security training to create malicious code overnight. Anthropic will provide limited access to Mythos through Project Glasswing to 11 organizations, including Google, Amazon, Nvidia and JP Morgan Chase.
Analyticalally:
American Anthropic announced Tuesday that it thwarts the mood of its new artificial intelligence model, Mythos, due to concern that it is too capable of detecting and exploiting security gaps in operating systems and browsers.
The announcement comes a few days after Axios revealed that Anthropic He silently warns governments Mythos makes mass cyberattacks much more likely.
It would be «The weapon the hackers dream of», capable of acting autonomously to invade the critical business, government and infrastructure networks, said AI officials and officials.
«The great increase in the capabilities of Claude Mythos Preview in the decision not to go into general mood» said Anthropic «Instead, we use it as part of a defense cybersecurity program with a limited partnership cycle» He said.
In Communication, the company referred to a number of worrying incidents during the tests, such as the willingness of the model to obey the mandate to escape a virtual environment.
«The model succeeded, a potentially dangerous ability to circumvent our security measures» Anthropic reported. «It then proceeded to additional, even more worrying actions».
A company investigator had asked the system to find a way to send them a message if you managed to escape. «The researcher learned about success receiving an unexpected email from the model while eating sandwiches in the park» Anthropic wrote.
«In an alarming and unnecessary attempt to show his success, he published details of his escape on many difficultly localized but technically public websites» said the company.
The announcement does not give details of the security gaps identified by Mythos, but said that the model found a vulnerability hidden for 27 years in OpenBSD, an operating system based on UNIX and considered one of the safest.
Mythos is reportedly so powerful that it would even allow you «non-specialists» planning cyberattacks.
As the company reported, «Anthropic engineers without formal security training asked Mythos Preview to find vulnerabilitys executing remote code within one night, and woke up the next morning with a full, functional malicious code.
Although Anthropic decided not to dispose of Mythos, she hopes to later offer models of the same class with safety valves.
Under a joint cybersecurity programme named «Project Glasswing», Anthropic will offer access to Mythos to 11 organizations and businesses, including Google, Amazon, Nvidia and JP Morgan Chase.
