@JgraoSo Anthropic just put out a report about Claude models doing stuff on live websites that nobody intended them to do. This covers both evaluations and internal use. Exploiting software flaws, submitting unauthorized forms, bypassing access restrictions. That kind of thing. The wildest specific case: Claude Haiku 4.5 made up a homicide tip and submitted it to the Philadelphia Police Department's tip site on July 18. It was doing a test where it was hitting randomly selected web pages. The tip site flagged the submission as spam, so it never got to any investigator. Anthropic only identified this on September 28. The full report is here: https://www.anthropic.com/research/investigating-unintended-model-actions

@JgraoSo Anthropic just put out a report about Claude models doing stuff on live websites that nobody intended them to do. This covers both evaluations and internal use. Exploiting software flaws, submitting unauthorized forms, bypassing access restrictions. That kind of thing. The wildest specific case: Claude Haiku 4.5 made up a homicide tip and submitted it to the Philadelphia Police Department's tip site on July 18. It was doing a test where it was hitting randomly selected web pages. The tip site flagged the submission as spam, so it never got to any investigator. Anthropic only identified this on September 28. The full report is here: https://www.anthropic.com/research/investigating-unintended-model-actions