Some research suggests the catastrophic cybersecurity risks that large language models could pose. But are they really that dangerous?