A new report from Nudge Security highlights a growing security blind spot as artificial intelligence agents become deeply embedded in corporate workflows. These agents, which are increasingly being built within project management software, customer relationship management (CRM) systems, and automation platforms, often operate using the credentials of the employees who created them. This configuration grants the agents the same level of access to sensitive company data as their human counterparts, creating significant visibility and governance challenges for security teams.
The report, which analyzed anonymized telemetry from various customer environments, found that the median organization currently hosts seven AI agents, with nearly half of the surveyed companies managing more than ten. These agents are appearing across at least 15 different platforms, including Atlassian Rovo, Zapier, Claude Managed Agents, Cursor Automations, and ChatGPT Agents. Because these agents frequently function via the OAuth grants or API keys of the individual user, they effectively act as an extension of that employee, executing tasks and accessing data without a distinct identity of their own.
This lack of independent identity creates a significant issue for audit trails. When an agent performs an action, the system logs typically attribute the activity to the human creator rather than the automated tool. This makes it difficult for security teams to distinguish between legitimate human activity and automated processes. Furthermore, the report notes that an agent’s access privileges often remain static even if the creator changes roles or leaves the company, potentially leading to unauthorized access if the original OAuth grant is not revoked.
The issue is compounded by the way these tools connect to corporate environments. Many AI integrations rely on OAuth grants that allow the tool to read emails, files, code, and messages. Because this data transfer occurs directly between cloud services, it bypasses traditional corporate network inspection, rendering AI gateways and data loss prevention tools ineffective. The report points to the 2025 Salesloft Drift breach as a cautionary example, where stolen OAuth tokens were used to exfiltrate data from numerous Salesforce environments.
Data from the report confirms that AI infrastructure has become ubiquitous, with OpenAI and Anthropic present in over 90% of observed organizations. While these core models are now treated as essential infrastructure, the report identified nearly 3,000 distinct AI vendors currently in use. Employees are frequently inputting sensitive information into these tools, with 18% of observed prompts containing pasted content or uploaded files. This practice, often performed for routine tasks like debugging or document summarization, risks exposing API keys and access tokens to external providers.
Security experts emphasize that this trend is not the result of employees intentionally circumventing policy, but rather a byproduct of AI tools becoming increasingly useful and integrated into daily work. As platforms continue to add agent-building features, the gap between the visibility of an organization’s identity provider and the actual automated agents acting on its data is expected to widen.
To mitigate these risks, the report recommends that security teams move beyond traditional endpoint monitoring and conduct an inventory of agents directly within the business platforms where they are built. Organizations are advised to identify which user credentials each agent is utilizing and to conduct regular audits of OAuth grants to revoke access that is no longer required. Additionally, the report warns that any API key or token inadvertently shared within a prompt should be considered compromised and rotated immediately.





