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