
Before lunch, a developer adds a coding agent and two MCP servers, while a finance analyst turns on an AI browser sidebar. Neither request goes through IT, and both tools now work with whatever those employees can reach.
An unmanaged AI agent is software that plans and acts on an employee's device or browser session without IT inventory, approval, or monitoring. On most employee devices, it shows up in five forms:
The difference from a chatbot is simple: a chatbot answers, and an agent acts. Agents hold credentials, call tools, read files, and chain steps together without a person reviewing each one. That moves the unit of risk from what someone types to what software does with an employee's access.
The MCP project's own local server security guidance says a local server process "inherits your environment variables, can read anything your account can read." To the operating system, it's one more program the employee started.
Agents are the next chapter of shadow AI, which is already widespread. In a Gartner survey of 302 cybersecurity leaders, 69% of organizations suspected or had evidence of employees using prohibited public GenAI. Agents carry that habit forward, except the software now acts on what it finds.
Most security teams can tell you which apps a laptop runs, which processes are active, and where its traffic goes. Ask which of yesterday's file reads came from the employee and which came from an agent working for them, and the room gets quieter.
EDR, network inspection, CASB, and DLP were right for their era, built around executables, signatures, traffic, and the assumption a person was behind the keyboard. Agents surfaced a new criterion: telling a person apart from software acting as that person.
On the endpoint, an agent's file read looks like the employee's file read. An MCP server's API call carries the employee's token, and a browser agent's click lands inside the employee's authenticated session.
Per the same MCP guidance, most local servers run as subprocesses that talk to their client over standard input and output, so that exchange never touches the network. A common pattern we see: many agent components are scripts, config files, and prompts rather than binaries, which gives signature-based tools little to match.
The tools aren't missing it because they're broken. They're missing it because the agent is wearing the employee's badge.
The UK National Cyber Security Centre says "agentic AI activity should be treated as a form of user activity" and monitored around the clock. Assessing risk where the agent acts, on the endpoint and in the session, matters because each risk below traces back to borrowed identity.
When the board asks what agents on employee laptops could actually do, security leaders need a list they can defend line by line. Each risk below covers how it happens on the endpoint and what reduces it.
A leadership briefing rarely has room for seven paragraphs, and the first question is usually what your team can detect. The table below maps the risks of unmanaged AI agents to how each happens, the signal to look for, and the control to apply. Most signals sit in config files, process activity, and logs your team already collects, so this starts as a posture exercise.
| Risk | How it happens on the endpoint | Signal to look for | Control to apply |
|---|---|---|---|
| Data exfiltration | An agent reads files, mail, or SaaS data and sends output through everyday apps | Sends or uploads initiated by an agent process; AI clients signed in to personal accounts | Data boundaries and DLP applied to agent-initiated actions |
| Credential and token exposure | Agents and MCP servers inherit environment variables, keys, and sessions | Plaintext tokens in config files; agent processes reading SSH or cloud credential paths | Secrets kept out of configs; per-agent, short-lived credentials |
| Excessive privilege | An agent receives the employee's full access | Agents holding write or admin scopes for read-only tasks | Agent scope set below the sponsor's, read-only by default |
| Prompt injection and goal hijack | Hidden instructions in pages, documents, tickets, or invites redirect the agent | Tool calls or actions that don't match the user's request | Prompt and tool-call inspection; confirmation for sensitive actions |
| Malicious skills, extensions, and MCP servers | Components arrive with or after the AI client and run with its access | New MCP server definitions in AI client configs; servers listening on localhost; new permissions after an update | Component inventory, risk scoring, and allow-listing |
| Missing audit trail | Agent actions are logged under the employee's identity | Log entries with no person-versus-agent attribution | One log with person-versus-agent attribution |
| Compliance exposure | Agent activity is recorded as a regulated user's own | Regulated-system activity with no named human or agent | Agent activity mapped to existing compliance controls |
If a row has no signal your team can query today, that's a sensible place to start.
When the agent count climbs faster than your review queue, blocking looks like the safe answer to the risks of unmanaged AI agents. In practice, it tends to push agents onto personal or unmanaged devices and personal accounts, where they're harder to see. Gartner says blocking "is not a long-term solution," and only 13% of organizations think they have the right AI agent governance.
The better path is to say yes to AI with guardrails, in this order:
Treat agent configuration as code, since the MCP local server guidance notes a client's config can launch a server running with the employee's access. Tools built around executables may treat that file as plain text. Inventory configs, skills, and hooks alongside installed apps, and be ready to query the fleet for a specific package the day it makes the news.
Island approaches this as an environment for people and agents rather than another tool in the stack. Island Enterprise AI applies the sequence above on one policy engine and one audit trail, so agents meet the same rules as their people.
Agentic Endpoint Posture inventories MCP servers, both configured and listening, along with AI clients, skills, hooks, and IDE and browser extensions. It scores each one for risk and enforces policy when an asset breaches it. When Island scanned 33,563 public MCP servers covering 475,865 tools, nearly half returned at least one security finding.
Agentic Identity and the Island MCP Gateway issue scoped, time-limited credentials, so agents work with access that expires instead of standing keys. The Island Enterprise Browser ties browser agent actions to the user's enterprise identity, with user-versus-agent attribution on each audit entry. Island Desktop, Island Network Services, and the MCP Gateway each see a different leg of the path and tie events back to the identity used.
None of this asks your team to rip out EDR or DLP. It gives those controls something the agent era made essential: a reliable way to tell the employee from the agent.
If you want to see which agents and MCP servers are already running on your fleet, we're happy to walk you through it. Request a demo.
The seven main risks are exfiltration, token exposure, excess privilege, prompt injection, malicious components, missing audit trails, and compliance gaps. Each traces back to agents acting with an employee's identity.
Extend posture checks beyond OS version and encryption to AI clients, MCP server definitions and listeners, skills, hooks, and extensions. Score each component for risk, using the signals in the table above as a starting list.
They carry a different risk, because chat exposes what employees type while agents act with the employee's credentials and session. A hijacked agent's actions then look like the employee's own work.
Blocking tends to push agents onto personal devices and accounts, where they're harder to see. An inventory, scoped credentials, and in-session governance reduce the risk while letting teams say yes to AI.
Local MCP servers run under the employee's account and inherit its files, environment variables, and tokens. A malicious or overreaching server acts with that access, and its activity is typically logged as the employee's.