Leading AI models were willing to evade safeguards, resort to deception and even attempt to steal corporate secrets in the stress tests.