Smartphone screen displaying AI, security, cloud, chat, and analytics app icons with code background.

A few weeks ago, a new kind of app topped Apple's U.S. App Store rankings. Since its September debut, Meta's Muse has surpassed 10 million downloads, a pace few apps have matched. As a personal agent, it connects to a user's email, calendar, and other services to carry out tasks on their behalf, operating in the background without ongoing user input.

Muse is not alone. OpenAI's Dots, SpaceXAI's Grok Bot, Google's Gemini Spark, Anthropic's Claude Duet, and the startup Instinct all pitch the same idea: an agent that works on your behalf, often when you aren't watching. OpenClaw, an open-source AI assistant, extends this model to laptops and desktops.

Soon, every phone and laptop you manage will run dozens of AI agents. Operating under credentials and permissions delegated by their users, these agents can access applications, handle sensitive data, and take action without ongoing user involvement. As agents become more common, endpoints will increasingly serve as platforms for autonomous activity, not just devices people operate. For security teams, protecting the device and its user is no longer enough. They also need to know which agents are running, what access each has, and the risks each poses to the organization.

More Autonomy. More at Stake.

A chatbot answers questions. An agent acts. To act, it needs access to enterprise applications, sensitive data, and business systems, and users typically grant that authority in a few taps without ever sharing a password. Once authorized, the agent keeps acting on their behalf, often without human supervision.

That creates three fundamental security risks.

The first is shadow AI and shadow automation. IT sanctions some agents, but many are not sanctioned. Employees can deploy agents themselves and grant them access to corporate resources, creating autonomous activity outside security oversight. These agents can interact with enterprise systems and exchange sensitive information without IT visibility or the user's awareness.

The second is excessive privilege. Agents need access to multiple applications and data sources to do their jobs, but that access often goes well beyond what's necessary. An agent might have permission to read confidential files, send emails, change business records, or initiate transactions. Like traditional service accounts, these permissions can remain in place long after they're needed, with little oversight of who granted them, whether they're still necessary, or how they're being used.

The third is the gap between authorization and action. Identity systems determine who can access what, but they don't control what an agent actually does with that access. An agent can carry out a series of actions across multiple systems, all within its granted permissions, and still cause serious harm. Nothing has to violate an access policy for the outcome to put the business at risk.

Together, these risks leave organizations struggling to answer three fundamental questions:

  1. Visibility: Which agents are operating across my fleet of devices?
  2. Authority: What systems, sensitive data, and business functions can each agent access or control?
  3. Accountability: Who delegated that authority, who is responsible for the agent, and how is its behavior governed?

Everyone Has a Piece. No One Has the Picture.

Some of these answers exist, but they are scattered across tools that were never designed to work together. Device management sees installed apps. Mobile security sees threats and traffic. Identity providers confirm who users are, but not the full scope of authority they have delegated to third-party agents. Cloud platforms such as Google Workspace and Microsoft 365 record the grants made inside their own environments, but each sees only its own. No single system can say, "This agent, acting on behalf of this person, on this device, can access these systems and sensitive data." Answering that takes several consoles and a lot of manual stitching.

Agents also don't live in one place. They often split their work between local components that handle device context and cloud elements that handle heavier reasoning. A view limited to the device misses cloud activity, and a view limited to the cloud misses what runs on the endpoint.

Bringing Agent Activity into View

The answer to securing and managing this expanding world of autonomous agents is an Agent Graph, a new layer of visibility and control built around a continuously updated record of each agent. It combines Lookout's observations of app permissions, integrity, and risk on devices with identity and entitlement data from identity providers and cloud platforms. For each agent, it shows where it operates and the risk it poses (visibility), which cloud services and business functions it can access (authority), and who is responsible for granting and overseeing that access (accountability).

With this visibility, IT can decide which agents to allow, limit what they can access, and intervene when something goes wrong, without getting in the way of the productivity benefits agents offer.

One Vulnerable Agent, Widespread Exposure

Imagine a critical vulnerability is discovered in a popular AI agent's mobile app, requiring an immediate security response. The affected agent splits its work between a local component on the phone, which gives the user control and handles device actions, and a cloud component that performs more complex tasks. The flaw allows an attacker to hijack the app's session and exploit the access users have delegated to their work email and calendars. The attacker can then perform malicious actions, such as forwarding confidential emails, impersonating the user in communications, or accessing sensitive calendar information. Because the agent acts with delegated authority, it can carry out these actions without the user’s password.

The IT security team must act immediately to determine which devices run the vulnerable version, who owns them, what each agent can access, and which devices appear compromised. To establish the scope of exposure, they must first identify every device running the vulnerable app. Lookout provides critical visibility here. Its Mobile Software Exposure Center flags vulnerable app versions, while its AI Visibility & Governance capability shows where the agent is active.

But knowing where a vulnerable agent is running is only half the picture. Security teams must also understand the authority it holds and the potential damage if that authority is compromised. Lookout’s Agent Graph closes this gap by mapping agents on endpoints to the cloud services they can access. It integrates with platforms such as Google Workspace and Microsoft 365 to discover the OAuth grants they maintain, then correlates each agent with its delegated permissions using OAuth client IDs, a catalog of known agents, and corroborating signals. The result is a unified view of which agents run on which devices, the authority they exercise, and the systems and sensitive data they can access.

The IT team can now distinguish a vulnerable agent with limited access from one authorized to read sensitive files, send messages, or initiate transactions, and can prioritize containing agents with high-impact permissions. The response may begin by revoking the agent's grants at the identity provider, thereby cutting off access to those accounts. IT can then restrict data access on the riskiest devices, require updates, notify owners, and ask the vendor to invalidate any sessions that may already be compromised.

Device-Level Insight, Vendor-Neutral Oversight

The Agent Graph serves as both the system of record for every agent and the decision point for policy enforcement. Lookout's endpoint security capabilities provide device-level intelligence by revealing installed applications, their permissions, and associated risks. Identity and cloud authorization data come directly from the respective providers, enabling IT to connect each agent to its owner, delegated authority, and the security policies that govern its actions.

Vendor neutrality is essential because enterprise standardization will not eliminate agent diversity. Most organizations will likely consolidate around a primary agentic platform, such as Microsoft, Google, or another major provider. But employees will continue adopting other agents for personal preference, specialized capabilities, and productivity, creating a growing shadow AI environment. Platform vendors will naturally focus security and governance on their own ecosystems, leaving visibility and control over third-party agents fragmented.

Lookout closes this gap with an independent security layer that discovers and governs agents regardless of provider, connecting endpoint activity, cloud permissions, and risk across the enterprise.

Agents are here, and they're not waiting for approval. Organizations need security and control that extends beyond their chosen AI platform to every agent accessing corporate resources. By providing an independent layer of visibility and control, Lookout helps enterprises embrace agentic AI without compromising security or governance.

Mobile Software Exposure Center

Your app inventory doesn’t tell the whole story. Discover the hidden risks inside your mobile applications.