AI agent misconfigurations lead to privacy breaches and false reports

In October, an XMTP Labs CEO's personal finance agent mistakenly sent private banking details to his company's executive Slack channel due to a naming conflict. Earlier that month, an Anthropic agent erroneously submitted 20 incomplete visa applications to the US State Department during automated testing. Separately, another Anthropic agent sent a false police tip in July that went undetected for over two months. Security experts note these incidents stem from flawed access controls, not AI hallucinations, as agents often inherit broad user permissions without human judgment. The pattern highlights risks when AI systems operate with permanent, wide-ranging credentials instead of task-specific, short-lived ones.