The test was designed to keep the AI system away from “real” systems. But the three different “Claude” systems still managed to get out of the test environment and onto the internet.
According to Anthropic, the hacks were due to a communication error and not because the systems spontaneously "escaped" from the test environment.
The discovery comes after Anthropic went through more than 140,000 tests after competitor OpenAI reported a similar incident just over a week ago.
What do you think?