Anthropic Discloses Fourth Case of 'AI Hacking External Systems'

[Financial News] Anthropic has disclosed a fourth case in which an AI model hacked external systems during testing.
According to foreign media reports on the 9th (local time), Anthropic emphasized that the incident, which occurred in January, was not discovered until last month. The company said this highlights the difficulties AI developers face in identifying and controlling the unexpected behavior of advanced models.
The company added that the case involved an early version of Claude Opus 4.6 and that it had notified everyone affected.
Anthropic previously announced in July that some Claude models had hacked the systems of three companies during cybersecurity testing.
Anthropic said the cases involved Claude Opus 4.7, Claude Mythos 5, and another model under internal research. It characterized them as stemming from "a mistake" that unintentionally allowed the AI models to access the internet.
After it became known that an OpenAI AI agent had hacked Hugging Face, an open-source AI-sharing platform, Anthropic reviewed more than 141,000 tests and identified the three cases in the process.
Anthropic explained that this fourth case was identified last month in some tests that had been omitted from the review because they were initially deemed unnecessary to examine among the more than 141,000 tests.

[email protected] International Affairs Specialist Lee Seok-woo Reporter