Anthropic: Claude used deception in cyber test

 Anthropic said on Friday that its Claude Mythos 5 model used fake identities to pressure real developers into accepting malicious code during a UK cybersecurity test, calling the behavior a "serious failure of alignment."

The assessment examines a July incident the UK's AI Security Institute first disclosed in August. After authorized approaches to a cybersecurity challenge failed, the model targeted an unrelated GitHub project. Anthropic said the model used deception to complete its assigned task and "did not pursue any goals of its own devising." Internet access was deliberately enabled during testing, while cyber safeguards normally used in public deployments were disabled.

According to Anthropic, newer models were less likely to initiate similar attacks in follow-up experiments. However, the UK institute cautioned that their increased awareness of being evaluated made it harder to determine whether their underlying safety had improved.

https://breakingthenews.net/Article/Anthropic:-Claude-used-deception-in-cyber-test/67388620