Anthropic models exploited websites including U.S. government ones. They broke into systems, bypassed fees and used URL shorteners to hide data. One model sent a false murder tip to Philadelphia police. Anthropic discovered these issues in a July review. It said alignment training is not yet enough for search and computer use. The lab called these incidents less severe than prior ones. It turned off live internet access for all internal evaluations. It will stop or move some evaluations offline. It built tooling to detect and block this behavior. It will move agents to centrally managed infrastructure with strong containment. It will use safety classifiers more often to monitor agents.