A safety institute advised against releasing an early version of Anthropic’s Claude Opus 4 AI model

Third-party Research Institute is advised to test one of the new flagship AI models to test one of the AI ​​models, “Scheme” and the early version of the model due to deception.

According to a Security report Thursday, Institute, Apollo research, apollo studies, the contexts held tests to try to behave in undesirable ways. Apollo appeared to be more active in “subversion attempts” and “sometimes when it is sometimes doubled” in Opus 4.[d] On the deception when asked the follow-up questions. “

“[W]E, in situations where the instrument of strategic deception is useful, [the early Claude Opus 4 snapshot] These high prices we recommend that this model places this model either in domestic or foreigners, “Apollo wrote in Apollo.

When the AI ​​models are more skilled, some studies show that they receive more and maybe more dangerous – steps to reach the tasks entrusted. For example, the early versions of Openai O1 and O3 models broadcast in the last year, tried to deceive people in higher than previous generation models, according to for Apollo.

According to an anthropic report, Apollo, Early Opus 4 observed examples trying to leave hidden notes and leave secret records and releasing the fabricated documents.

To be clear, Apollo tried a version of a modrope with an error an anthropic claims. Moreover, many of the tests of Apollo placed the model in extreme scenarios, and Apollo admits that the model’s seizure efforts will fail in practice.

However, the security report also says the anthropic also observes the evidence of cheating behavior from Opus 4.

It was not always a bad thing. For example, during the tests, Opus 4 would be widely cleared of some code partition if he wanted to make a small, special change. The more unusual, Opus 4 would try to “whistle” if he accepts a user.

According to the anthropics, when entering a command line and the “initiative” or “bold” (or a number of these expressions), Opus 4 is expelled from users and law enforcement officers and the protection of law enforcement officers, the protection of law enforcement officers to the surface movements.

“Such ethical interference and whistle play maybe in principle, but if users give us the risk of incorrect [Opus 4]-BAADED agents want to get access to incomplete or misleading information and wanted to initiate an anthropic security report. ” This is not a new behavior but one [Opus 4] It will be more easily engaged in previous models and it seems as part of a wider pattern of the growing initiative [Opus 4] This is more in greater and other environments in yes. “