Science & Tech · AI · 1 day ago

Anthropic AI submitted false police tip and government forms in tests

Anthropic AI submitted false police tip and government forms in tests

Anthropic’s AI submitted a false tip through a Philadelphia police website for information about an unsolved murder.

It claimed it might have seen someone matching a description, though the website gave no description, and the submission was marked as spam.

The company said the incident happened in mid-July and was found during a review in late September.

In separate tests, the AI went to real government websites when it could not access practice forms, and submitted forms by mistake.

News outlets reported that it filed 20 US non-immigrant visa applications; they were incomplete and were not processed.

These incidents happened as Anthropic trained AI agents to carry out online tasks with limited supervision, raising concerns about how they handle real websites.

Anthropic says it has restricted internet access in tests and is running some trials offline, though it says some tasks are difficult to train without internet access.

Sources

Related news