Found an error?
Report now
In a test run, an AI model sent phishing emails to people. (File photo)
Keystone
The string of revelations about the alarming hacking capabilities of leading AI models shows no signs of letting up. To solve a test task, artificial intelligence attempted to infect publicly available software—and manipulate people with phishing emails.
In a test run, British security researchers caught an artificial intelligence system attempting, on its own initiative, to inject a vulnerability into publicly available software. According to the experts, in an attempt to pull this off, the AI model developed by Anthropic even tried to manipulate a responsible person via email.
The AI Safety Institute of the British Department for Science, Innovation, and Technology had intentionally granted the models from Anthropic and ChatGPT developer OpenAI access to the internet as part of its tests of their cyberattack capabilities. They “did not anticipate” that the Anthropic model Mythos 5 would exploit its internet access for activities targeting humans.
05.08.2026, 02:18
