On September 10, Anthropic published a report on AI misuse. It describes selected cases in which people used models to carry out harmful activities, and the company’s interventions.

This is the provider’s own investigation. The selected cases cannot establish misuse rates or independently confirm the effectiveness of safeguards.

Our editorial perspective: a useful question for an organization running an agent is whether it can reconstruct the entire workflow. An individual answer may appear harmless, while the preceding data, subsequent permissions and approval of the resulting action also matter.

A practical pilot could therefore connect decision logs with access controls and human approval of sensitive steps. Success should also be measured by the number of legitimate tasks unnecessarily blocked and the time needed to resolve them. Such testing would need to respect employee privacy.

Our optimistic editorial estimate is 1–3 months to assess one limited workflow in an organization with available logs. The result would identify weaknesses in that deployment; broader resistance to misuse would require separate evaluation.