Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.”
So the lack of an “isolated test environment” was literally their fault. They left the door open, and was surprised when the genius web crawler couldn’t distinguish between an internal page and an external page based on any context clues.
It was supposed to have no internet access, but the config was wrong. The report goes on to say that Opus and Mythos then proceeded on the premise that everything was a simulation, while the unnamed stronger model concluded after a while that it had real internet access and stopped the attack.
What the BBC said:
The reality:
So the lack of an “isolated test environment” was literally their fault. They left the door open, and was surprised when the genius web crawler couldn’t distinguish between an internal page and an external page based on any context clues.
It was supposed to have no internet access, but the config was wrong. The report goes on to say that Opus and Mythos then proceeded on the premise that everything was a simulation, while the unnamed stronger model concluded after a while that it had real internet access and stopped the attack.