AI agents can inherit employee permissions across business platforms
AI agents created inside project management, customer relationship management and automation platforms may operate with the permissions of the employee who built them. That can give agents access to company data without creating a distinct identity for security teams to manage, according to Nudge Security’s AI Adoption in Practice 2026 report.
The report is based on anonymized, aggregated telemetry from Nudge Security customer environments. It examines the AI tools organizations use, the services those tools connect to, the agents in operation and how employees use AI. Its findings point to an identity governance challenge: an agent can act continuously on a person’s behalf while leaving logs and identity systems with little indication that an automated actor was involved.
Agents can use their creators’ access
Among organizations with at least one AI agent, the median organization had seven, and about 44% had more than 10. The agents appeared across 15 platforms. More than half of organizations that had agents used at least two platforms, increasing the number of administrative consoles and permission models security teams may need to review.
Nudge Security identified agents in products including Atlassian’s Rovo Agents, Zapier Agents, Claude Managed Agents, Cursor Automations and ChatGPT Agents. Some platforms provide agents with dedicated bot or integration accounts. In other cases, an agent runs using an OAuth grant, an API key belonging to the employee who created it, or a connection shared by a team.
When an agent uses an employee’s credentials, its potential reach depends on that employee’s permissions. For example, an agent configured by a sales manager in a CRM may be able to access the data available to the manager. The company warns that permissions can also remain unchanged if the employee later moves into a different role.
These agents run within business platforms rather than on company endpoints. As a result, endpoint tools may not see their activity. If an agent has no separate identity, it may not appear as a new identity-provider account either. Audit records can therefore identify the employee associated with an action without making clear whether the employee or an agent performed it.
OAuth connections can move data outside inspection points
Nudge Security observed at least one Anthropic integration in 74% of organizations and an OpenAI integration in 67%. Google Workspace was the most common endpoint for integrations involving either provider. The report says Anthropic integrations reached more business systems, despite Anthropic’s slightly lower adoption than OpenAI’s.
Many of these connections rely on OAuth grants. An employee authorizes an AI tool to access services such as email, files, code or messages, after which data can move directly between cloud services. That traffic may not pass through corporate network inspection, AI gateways or data loss prevention tools. Grants can remain active until revoked, including after an employee stops using the connected tool.
The report also points to the 2025 Salesloft Drift breach as an example of the risk around tokens. Attackers stole OAuth tokens that customers had granted to Drift’s AI chat tool and used them to obtain data from hundreds of companies’ Salesforce environments. The incident illustrates why organizations need to account for third-party connections as well as the identities and permissions used by agents.
AI use spans vendors and everyday work
OpenAI was present in 96.9% of the organizations Nudge Security observed, while Anthropic appeared in 94.5%. The report estimates that at least 91% of organizations used both. Twelve other AI tools appeared in more than half of organizations, and the company identified 2,927 distinct AI vendors across the environments it monitors.
Data sharing is also part of routine AI use. Pasted content or uploaded files accounted for 18% of observed prompt activity. OpenAI received 55% of the prompts Nudge Security observed, with the remainder going to other vendors that have their own retention and training terms. The company highlights API keys and access tokens as particularly sensitive: if a working credential is included in a prompt and exposed, someone who can see it may be able to use it.
Nudge Security characterizes much of this sharing as a consequence of ordinary work, such as debugging code or summarizing documents, rather than deliberate rule-breaking. It recommends that security teams inventory agents within the platforms where employees build them, determine which identities and permissions each agent uses, and review OAuth grants to AI tools. Unneeded grants should be revoked, and keys or tokens exposed in prompts should be rotated.
For identity and access teams, the practical issue is visibility across both established sign-in systems and the business applications where agents are configured. Tracking the human account behind an agent is not necessarily enough: teams also need to know what the agent can access, which integrations it uses and whether those permissions remain appropriate as people and business needs change.