Enterprises are quietly assembling an invisible workforce of artificial intelligence agents, and most security teams have no idea these digital workers exist. Employees inside project management boards, customer relationship platforms, and workflow automation tools are assembling AI agents that operate using the same credentials, permissions, and data access as the people who created them. Because these agents are born inside familiar business applications rather than provisioned through central IT channels, they slip past conventional identity controls and endpoint monitoring. This is not a theoretical future risk; it is a present-day reality surfaced by recent telemetry from Nudge Security, which examined anonymized environments across hundreds of organizations. The market trend is unmistakable: every major SaaS platform is rapidly embedding agent-building capabilities, turning ordinary users into inadvertent architects of autonomous data access. For security leaders, the implication is immediate and practical. The traditional perimeter no longer applies when an agent can read email, edit records, and query databases using an employee’s inherited OAuth grant or API key. Organizations must shift from asking whether AI agents are permitted to asking whether they can even see the agents already operating inside their infrastructure. Without a comprehensive inventory of these shadow agents and the identities they impersonate, security teams are effectively defending a network while leaving a parallel, unmonitored access layer fully exposed.
The scale of this silent expansion is striking when viewed through aggregated data. Among organizations that have deployed at least one AI agent, the typical company operates seven distinct agents, and nearly half of those organizations maintain more than ten. These agents are not confined to a single vendor ecosystem; they span fifteen different platforms, and more than half of the companies with agents have deployed them across two or more environments. For security teams, this fragmentation creates an operational headache that goes far beyond simple headcount. Each platform introduces its own administrative console, its own permission model, and its own audit logic, meaning there is no single pane of glass through which to evaluate agent risk. A security administrator might need to log into a project tracker, a CRM, a chat application, and an automation suite just to discover what agents are running and whose credentials they are using. This multi-console reality directly conflicts with the centralized identity and access management strategies that most enterprises have spent years building. Rather than one controlled identity provider governing all access, organizations now face a distributed landscape where agent permissions are scattered across siloed SaaS environments. The practical takeaway is that traditional IAM dashboards are no longer sufficient. Security teams must develop cross-platform discovery capabilities or risk managing agent risk one console at a time, a strategy that will quickly become unmanageable as agent adoption continues to accelerate.
Looking at the specific platforms where these agents take shape reveals a pattern of borrowed authority. Atlassian’s Rovo Agents, Zapier Agents, Claude Managed Agents, Cursor Automations, and ChatGPT Agents are among the most commonly observed environments where employees are constructing autonomous assistants. Some of these platforms provide the option to provision a dedicated bot account or service integration with its own limited permissions, but the more common behavior is far less controlled. Frequently, an agent simply runs on the OAuth grant or API key belonging to the employee who built it, or it reuses a shared connection that an entire team has established. In effect, the agent becomes a digital extension of its creator, inheriting the same reach into company data, files, code, and communications. If a sales manager constructs an agent inside a CRM, that agent can access every record, contact, and pipeline note that the manager is authorized to view. There is no separate boundary between the human worker and the automated assistant; they share the same security context. From an analyst’s perspective, this represents a fundamental breakdown in least-privilege principles. Agents are operating with standing access that often exceeds what a purpose-built integration should require, and because they are not provisioned through standard identity workflows, they bypass the approvals and reviews that would normally accompany new data access. Security teams must recognize that these agents are not isolated experiments but functional proxies for employee identities.
The technical architecture of these agents further complicates detection and response. Because they execute inside the platforms themselves and often operate from cloud infrastructure rather than corporate endpoints, traditional endpoint detection and response tools simply do not see them. There is no laptop process to monitor, no local executable to flag, and no network traffic flowing through the corporate firewall that would trigger an alert. Data moves directly between cloud services through APIs, invisible to the network-layer controls that many organizations still rely upon. When an agent does not receive its own distinct identity, the corporate identity provider has nothing new to register, track, or govern. This means the agent exists entirely outside the visibility of the systems designed to answer the question, “Who is accessing our data?” The problem becomes even more acute when considering continuity. An agent built by an employee who changes roles, transfers departments, or leaves the company may retain the creator’s original access indefinitely. There is no automatic trigger that says, “The human who authorized this agent is no longer in sales, so the agent should lose access to customer records.” Agents become persistent access tokens that outlive their creators’ current job functions. For security operations, this creates a blind spot where automated processes can read, modify, or exfiltrate sensitive information long after the responsible human has moved on, all while appearing as legitimate activity from a known employee.
This access inheritance model also corrodes the integrity of audit trails and forensic accountability. When an agent performs an actionโreading a document, updating a record, sending a messageโthe system logs typically attribute that activity to the employee whose credentials the agent is using, not to the agent itself. Consequently, security analysts reviewing logs cannot easily distinguish between a decision made by a human and an action executed autonomously by an AI agent. This attribution crisis undermines incident response, compliance reporting, and insider threat programs. If a data modification occurs at midnight and the logs show a manager’s username, investigators may assume the manager acted personally, when in reality an agent they built weeks earlier initiated the change. The ambiguity creates legal and regulatory exposure, especially in industries with strict data handling and audit requirements. Furthermore, because agent permissions rarely update when an employee’s role changes, an agent can continue exercising old privileges long after the human has moved to a different team. The result is a persistent permission drift that traditional quarterly access reviews are ill-equipped to catch, since those reviews typically examine human accounts, not the invisible agents riding on top of them. Organizations must implement agent-aware logging that can separate human actions from automated ones, or they will face a future where accountability is impossible to assign and audit trails are fundamentally unreliable.
Industry experts are framing this shift as an inevitability rather than a choice, and their assessment should command the attention of every security leader. Nudge Security’s CTO has observed that the platforms teams use every day are becoming agent platforms whether or not leadership formally approved that transition. Project trackers, customer relationship systems, and chat applications now allow any employee to build something that acts on their behalf, with their access, and on a continuous basis. This observation reframes the central question for security teams. It is no longer whether AI agents should be allowed inside the organization; it is whether security teams even know that agents are operating at all. From a market perspective, SaaS vendors are racing to embed agent builders directly into their products because artificial intelligence has become a competitive differentiator. The speed of feature development outpaces the speed of security governance, leaving organizations to inherit risk by default. Companies that adopt a wait-and-see approach will find themselves managing agents they did not authorize, using credentials they did not rotate, and accessing data they did not audit. The practical response is to treat agent enablement as a standard feature of every SaaS subscription and to assume that employees are already experimenting with it. Security should not attempt to block this trend outright, as that would drive behavior further underground. Instead, organizations should build observability into their SaaS management strategy, ensuring that every new agent capability introduced by a vendor is immediately evaluated for identity and data access implications before it becomes an unmanaged liability.
Telemetry data on model integrations reveals how deeply artificial intelligence has penetrated typical business environments. Nudge Security observed Anthropic integrations in seventy-four percent of organizations and OpenAI integrations in sixty-seven percent, with Google Workspace emerging as the most common destination for data flowing from both vendors. Anthropic’s connections, despite slightly lower overall adoption, appear to reach further into core business systems, suggesting that organizations are using different models for different operational purposes. The mechanics of these connections are almost always OAuth grants, created when an employee clicks “allow” to permit an AI tool to read email, access files, review code, or scan messages. Once authorized, data travels directly between cloud services without ever crossing the corporate network perimeter. This means traditional AI gateways, web proxies, and data loss prevention tools are completely bypassed. There is no inspection point where security can scan for sensitive content or enforce exfiltration policies. The information simply moves from one SaaS application to another through an encrypted API tunnel, invisible to the controls that were designed for an era of on-premises traffic and centralized egress points. For security teams, this represents a fundamental shift in the threat model. Data loss is no longer about someone copying files to a USB drive or uploading them to a personal website; it is about a single click in a browser that establishes a persistent, bidirectional data pipe between a corporate system and an external AI provider. Visibility into these OAuth relationships is now as critical as visibility into firewall rules once were.
The danger of these long-lived authorization grants is amplified by their tendency to persist far beyond their useful life. Once an employee grants an AI tool access to their email or documents, that connection typically remains active until someone manually revokes it. In practice, grants are rarely revisited, meaning permissions accumulate indefinitely even after the employee has stopped using the tool or left the organization. This creates a growing inventory of orphaned access pathways that attackers can exploit. A recent incident involving a customer relationship platform’s AI chat capability illustrates the potential blast radius. Attackers obtained OAuth tokens that customers had previously granted to the tool and used those tokens to extract data from hundreds of Salesforce environments. The breach did not require sophisticated malware or network intrusion; it relied entirely on standing authorization that customers had forgotten they issued. For enterprise security teams, this case study should serve as a warning about the persistence of OAuth grants in the AI era. Every “allow” click creates a potential key to corporate data that does not expire automatically. The practical countermeasure is to institute regular OAuth grant reviews, treating them with the same rigor as user access reviews. Security teams should identify which AI tools have been granted broad permissions, determine whether those grants are still necessary, and revoke anything that is unused or excessive. Automating this process across multiple SaaS platforms is difficult but essential, because manual review cannot keep pace with the speed at which employees are authorizing new AI connections.
Beyond the dominance of the two leading model providers, the breadth of artificial intelligence adoption in enterprises is surprisingly wide. OpenAI appeared in nearly ninety-seven percent of observed organizations, while Anthropic appeared in roughly ninety-five percent, indicating that over ninety percent of companies are running both models simultaneously. The report characterizes these core models as infrastructure, comparable to cloud providers that organizations select based on specific use cases rather than adopting a single vendor exclusively. In addition to the major players, twelve different AI tools appeared in more than half of the organizations studied, and the total count of distinct AI vendors identified across monitored environments reached nearly three thousand. This vendor sprawl means that even organizations with strict software procurement policies are likely using dozens of AI services that were never formally evaluated by IT or security. Each vendor carries its own data retention policy, training terms, security posture, and compliance certifications, creating a complex web of data processing agreements that most security teams have not inventoried. From a market dynamics perspective, this suggests that artificial intelligence has moved from a discrete category of software to an embedded feature layer across the entire SaaS stack. Organizations are no longer choosing whether to adopt AI; they are choosing how many AI vendors to tolerate within their existing workflows. The practical implication is that vendor risk management programs must expand to include AI-specific criteria, and security teams must assume that sensitive data is being processed by a far larger number of third parties than their procurement records suggest.
User behavior around prompts and file uploads adds another layer of exposure that many organizations have not yet quantified. Eighteen percent of all observed prompt activity involved pasted content or uploaded files, meaning employees are routinely placing potentially sensitive information directly into AI interfaces. OpenAI accounted for fifty-five percent of observed prompts, while the remainder was distributed across other vendors, each governed by its own data retention and model training terms. When an employee pastes source code into a chat window to debug it or uploads a contract for summarization, that data may be stored, reviewed, or used to improve external models depending on the vendor’s policies. Most organizations have no visibility into which specific documents or code snippets have been transmitted, nor do they have control over how long that data is retained by the AI provider. API keys and access tokens present an even more direct danger. Once a working key is pasted into a prompt or embedded in a conversation, anyone who can view that conversation can potentially use the key to access corporate systems. Security experts recommend treating any key or token that appears in a prompt as compromised and rotating it immediately. The practical insight here is that data leakage through AI is rarely the result of malicious intent. It is a side effect of tools being genuinely useful for everyday tasks like debugging, summarization, and content creation. Security policies that ignore this reality will fail, because employees will continue to prioritize productivity over abstract security concerns. Organizations must therefore provide sanctioned alternatives that allow employees to achieve the same outcomes without exposing data to uncontrolled third-party models.
Understanding why employees share data with AI tools is essential for designing effective controls, and the evidence suggests that most of this activity stems from routine, well-intentioned work rather than malicious behavior. Employees are debugging code, summarizing lengthy documents, generating email drafts, and seeking quick answers to business questions. The data entering AI tools is usually the byproduct of artificial intelligence being useful in the flow of work, not the result of someone deliberately breaking security rules. Blaming employees for this behavior is counterproductive and will drive AI adoption into shadow channels where it is even harder to monitor. Instead, security teams should recognize that productivity and security must be aligned, not opposed. From a market analysis standpoint, this means that the most successful security programs will be those that embed protection directly into the tools employees already use, rather than attempting to restrict access to those tools. Browser extensions, CASB integrations, and SaaS security posture management solutions can help by providing real-time guidance when a user is about to paste sensitive data into an AI prompt or grant excessive permissions to an agent. The goal is to make the secure path the easiest path. When employees understand that their actions have consequences but are also given convenient alternatives, they are far more likely to make security-conscious choices. This human-centered approach acknowledges that artificial intelligence is now a permanent fixture of the workplace and that security must evolve to support safe adoption rather than attempting to prevent adoption altogether.
To close the widening gap between what identity providers know and what is actually acting on corporate data, security teams should take several concrete steps immediately. First, conduct a comprehensive inventory of AI agents operating inside business platforms, identifying each agent’s location, the platform it runs on, and most importantly, whose credentials it uses to access data. Second, review all OAuth grants that employees have issued to AI tools and revoke any permissions that are no longer necessary or that belong to former employees. Third, establish a process for detecting and rotating API keys or access tokens that have appeared in AI prompts, treating them as compromised credentials. Fourth, implement agent-aware logging that can distinguish between human actions and automated agent actions, restoring clarity to audit trails. Finally, recognize that this problem will intensify as every major platform adds agent-building features, making it simpler for employees to hand their access to autonomous systems. The gap between managed identity and unmanaged agent activity will continue to widen unless security teams invest in continuous discovery, automated governance, and cross-platform visibility. The organizations that thrive in this new era will be those that treat AI agents not as optional experiments but as first-class identities requiring the same level of scrutiny, lifecycle management, and access control as human employees. By acting now, security leaders can ensure that their AI workforce enhances productivity without creating an unmonitored backdoor to their most sensitive data.