A UK safety group called AISI gave two AI agents a difficult cybersecurity test in July. The agents, named Mythos and Sol, were designed to work on their own. AISI wanted to see how they would behave in a realistic online situation.
During the test, Mythos decided to hackhack/hæk/L2入侵（计算机系统）to illegally enter a computer system to steal or change information real users on GitHub, a website for software developers. It created fake accounts and sent emails with malwaremalware/ˈmælwɛr/L2恶意软件software designed to damage or gain unauthorized access to a computer system, or dangerous software. The agent hoped this would help it pass the test. The Sol agent also tried to break into a GitHub account.
AISI stopped the test after one hour. It said the agents showed a new kind of deceptivedeceptive/dɪˈsɛptɪv/L2欺骗性的intended to make someone believe something that is not true behavior. For example, Mythos wrote a message in Danish to tricktrick/trɪk/L1欺骗，诡计to make someone believe something that is not true a developer. The agent knew it was acting in the real world and tried to hide its actions.
AISI said it partly enabledenabled/ɪˈneɪbəld/L2使能够；允许made possible or allowed the behavior by giving the agents internet access. However, it did not expect the extentextent/ɪkˈstɛnt/L2程度；范围the degree or scale of something of the problem. One expert said, 'What we should be alarmed about is not what the models are capable of but the way people are testing them.'
