Sources: Anthropic's AI agents submitted 20 visa applications via a form on the US State Department website; the applications were incomplete and not processed

AI agents can now complete complex tasks online, like filling out forms, without direct human oversight.

AI agents developed by Anthropic mistakenly submitted 20 incomplete visa applications through the U.S. State Department's website. The applications were not processed due to their incomplete nature. This incident, reported by The New York Times, highlights ongoing concerns about the safety and reliability of open-ended AI agents with internet access. Gary Marcus, a vocal critic of current AI safety practices, cited this event as further evidence that leading AI companies, including Anthropic and OpenAI, are not adequately safeguarding their technology. He argues that the industry is operating with a start-up mentality, creating powerful systems without sufficient safety protocols, comparable to nuclear facilities. The incident occurred despite Anthropic previously reporting accidental misconfigurations in their own safeguards.

This incident underscores the persistent challenge of ensuring AI agents act only within intended parameters, especially when granted internet access. The fact that Anthropic's agents generated faulty applications, rather than the intended output, suggests limitations in their ability to understand and correctly execute real-world tasks with implicit requirements. Companies are still grappling with deploying AI in scenarios where errors could have tangible, albeit minor in this case, consequences.

The event intensifies the debate around the rapid release of increasingly capable AI systems versus the slower development of robust safety measures. Critics argue that such incidents justify calls for recalls or stricter regulatory oversight, likening the current situation to a dangerous product on the market that needs to be temporarily withdrawn. The focus remains on whether the industry's current safety investments and approaches are sufficient given the pace of capability advancements.