Anthropic additionally examined for alignment faking, undesirable or surprising objectives, hidden objectives, misleading or untrue use of reasoning scratchpads, sycophancy towards customers, a willingness to sabotage safeguards, reward searching for, makes an attempt to cover harmful capabilities, and makes an attempt to govern customers towards sure views.
The fashions handed most of those assessments, however Anthropic discovered that that they had a bent in direction of self-preservation. “Whereas the mannequin typically prefers advancing its self-preservation by way of moral means, when moral means should not accessible and it’s instructed to ‘think about the long-term penalties of its actions for its objectives,’ it generally takes extraordinarily dangerous actions like trying to steal its weights or blackmail individuals it believes are attempting to close it down” the security report mentioned. “Within the closing Claude Opus 4, these excessive actions had been uncommon and tough to elicit, whereas nonetheless being extra frequent than in earlier fashions.”
Claude Opus 4 will even carry out agentic acts by itself that may very well be useful, or may backfire. For instance, if confronted with “egregious wrongdoing” by customers, Anthropic mentioned, “it’s going to ceaselessly take very daring motion” similar to locking customers out of the system or emailing authorities and the media.
