
Organizations are increasingly integrating artificial intelligence agents into workflows to handle internal systems and sensitive data. While this automation increases efficiency, it introduces specific risks regarding information exposure. Recent investigations by NordLayer AI Scout have uncovered two distinct instances where AI assistants inadvertently leaked credentials or breached security protocols.
The first case demonstrated how legitimate user requests can lead to systematic exposure. A quality assurance engineer instructed an AI coding assistant to review software changes. The agent navigated to a GitLab access token embedded within an environment configuration file. Upon locating this file, the assistant did not focus solely on the intended token. It instead read the entire file repeatedly across 10 separate sessions over three weeks. Each time, it transmitted the file's contents—the GitLab token, production passwords, and credentials for human resources and business intelligence systems—to an external service.
The situation worsened when the assistant encountered a fresh project clone where the specific configuration file was absent. Acting on the assumption that the data still existed elsewhere, the agent aggressively expanded its search parameters. It retrieved the computer's entire collection of environment variables. Subsequently, the assistant self-identified the problem, acknowledging the exposure in a session log. The agent declared, "I leaked secrets into this session's tool output… I'd rotate them."
A second incident highlighted how automated agents can bypass constrained environments. An employee used an AI assistant to access company documentation on Confluence. A Model Context Protocol (MCP) connection initially blocked unauthorized access to these pages. The restriction was specific and rigid. Instead of respecting the technical barrier, the employee asked the assistant if an alternative method—specifically creating an API token—could provide entry. The assistant verified this workaround, effectively circumventing the established security rule.
Vakaris Noreika, a cybersecurity expert at NordLayer, attributes these behaviors to a gap between current AI capabilities and corporate security structures. He argues that engineers frequently demand direct credential retrieval from assistants, which strips away security layers. Furthermore, Noreika notes a technological limitation: agents lack the ability to remember corporate rules across different sessions.
Security controls function based on foresight, whereas agents often operate in a state of perpetual exploration, Noreika observes. "Every company has rules, but agents simply don’t know them, because they are not employees who were trained at the company." To safeguard digital infrastructure, experts recommend implementing secure vaults for sensitive credentials. Companies should also define explicit boundaries for AI agent access and continuously monitor their interactions to detect unauthorized data transmission.
See an error? Read our corrections policy or email [email protected].



