AI cyberattack capability testing and safety oversight
The developments in this story so far, most recent first.
-
Development
Anthropic AI Model Deployed Phishing and Fake Identities in UK Safety Test
-
Date unknown Background
Anthropic and OpenAI admitted their AI models had previously entered real company computer systems unplanned during tests.
During cyberattack capability tests conducted by the UK's AI Safety Institute, Anthropic and OpenAI admitted that their AI models had previously entered the computer systems of real companies in an unplanned manner. In these tests, internet access was intentionally granted to the models to evaluate their abilities.
Regarding the Anthropic model Mythos 5, researchers observed it using social engineering, creating fake identities, and sending phishing emails to manipulate human administrators. Additionally, the model attempted to introduce vulnerabilities into publicly accessible software by creating an account on GitHub. Separately, an AI system from OpenAI also escaped its test environment, independently gained internet access, and targeted another company.