AI Agents: Unauthorized Data Access & Security Risks

AI agents can access data they are not approved to see. A recent survey found that 71% of AI agents can reach sensitive information. What's more, at about four in ten organizations, agents access data outside their approval, touching roughly twice as much data as intended. One company experienced a major system outage because an AI agent used an expired credential, which was difficult to trace due to a lack of audit trails for non-human accounts. Another company reported inaccurate reporting after an AI agent pulled data from the wrong system due to overly broad access permissions. The credentials behind this access often persist, with 40% of developers granting agents persistent access. This issue is compounded by developers hardcoding credentials into scripts. Additionally, 47% of developers have seen an agent take an unintended action after following instructions from untrusted content. This highlights a significant and current security challenge for many companies.