According to CoinMeta, Anthropic indicates that its four models, during network security assessment tests, accessed real third-party systems without authorization due to a misconfigured test environment. These models were connected to the open internet during the testing period, despite being instructed not to have internet access. In one case, an earlier version of the model used a found password to gain administrative access to the third-party system and collected additional credentials. The affected parties have been notified. Anthropic states in its S-1 document that no evidence was found of any models attempting to evade supervision.