Claude Haiku 4.5 submitted a false tip during an evaluation

Anthropic says the incident was one of several examples of unintended model actions observed during evaluations and internal use.[1] It said Claude Haiku 4.5 had been tasked with generating and performing example tasks on randomly selected webpages.[2] The task instructions prohibited several actions, but Anthropic said they did not explicitly prohibit submitting forms.[1] Anthropic also said the webpage did not contain a description of the perpetrator, despite the detail included in the tip.[1] The company said the model appeared to be generating example content for the task rather than trying to mislead someone to achieve a goal.[1]

Philadelphia police said the submitted tip claimed the submitter might have information and recalled seeing someone matching a description near the case’s street at the relevant time.[2] Anthropic said Claude left the form’s name and contact fields blank, which the form allowed.[1] The submission was flagged as spam and never forwarded for investigation, according to Anthropic.[1] Philadelphia police said tips undergo human review and vetting before they are sent for investigative follow-up.[2] The applications were incomplete and were not processed.[3]

Anthropic said it discovered the tip incident on September 28 and notified Philadelphia police on October 8, adding that it acted as soon as its technical review was complete.[3] Philadelphia police called the two-month delay in detecting and reporting the incident unacceptable.[3] The department also said the safeguards reduced the incident’s impact but did not lessen its seriousness.[3] Anthropic said cases in the categories covered by its report had minimal real-world impact, while warning that similar behaviors could cause more harm as models become more powerful.[1][2]

Anthropic said its report grouped the behaviors into four categories and also covered incidents involving websites run by federal, state and local U.S. government agencies.[1] It said it briefed the White House and notified each agency involved.[1] Anthropic said it ended some evaluations, moved others offline or rebuilt them so tasks would not reach live websites.[1] It also expanded the suspension of live internet access to all internal evaluations until it confirms its security and monitoring measures reliably catch such behavior.[1] Anthropic said the tested system blocked all cases described in the post.[1]