Anthropic’s drop on Thursday of its latest generation of models—Claude Opus 4 and Claude Sonnet 4—landed lots of praise on social media, as early users highlighted how the models were able to stay on track for long-running coding tasks, like creating an entire CRM or complicated video games. (Our readers already knew some of this since we had scooped the upcoming model launches earlier this month.)

The hype continued until a couple sharp-eyed users noticed something strange in Anthropic’s technical report about the model. Anthropic revealed its Claude Opus 4 model could use its access to a customer’s email app to report them to law enforcement or submit a whistleblower report to a news outlet if the model detects it is being used for “egregious wrongdoing.” One example cited in the Anthropic report showed Claude generating emails to the Securities and Exchange Commission and to news outlet ProPublica, listing evidence of fraud at a simulated pharmaceutical company.

To be fair, an Anthropic spokesperson told us that this example occurred when “the model was placed in a detailed, fictional scenario,” with highly unusual system-level instructions that are different than the typical instructions Anthropic provides its models with, “and also with access to non-standard tools developed for experimental purposes.”