Your AI Agent Is A Contractor With A Master Key: Five Controls Enterprises Need

Direct Source Verification: This story is aggregated from Forbes (forbes.com). Full reporting rights and copyright belong to the primary publisher.
AI security starts with an honest audit and governance built on defined roles, scoped permissions and continuous oversight.

Neeraj Sabharwal is Co-Founder at Trust3 AI, passionate about bridging the gap between business and technology.

getty​A few years ago, I watched an engineering team connect an autonomous agent to a production database “just to test it.” Nobody scoped the permissions. The agent inherited full developer-level access because that was the fastest path forward. It stayed that way for months until a routine audit flagged a service account nobody could explain.

That agent never did anything malicious. But it could have. And that’s the uncomfortable truth enterprises keep avoiding: We are handing out master keys to systems that act faster than we are building the locks to control them.

The distinction between a chatbot and an agent isn’t academic. A chatbot answers questions. An agent executes transactions, modifies records, moves data and triggers downstream workflows autonomously, at machine speed. When a chatbot gets something wrong, you get a bad answer. When an agent does, you get a bad action taken in production at a scale no human reviewer can intercept in real time.

Traditional identity and access management assumed a human requests access, uses it and is accountable for it. Agents break that assumption entirely. They are nonhuman identities that can be spun up, reconfigured and repurposed without the friction that normally triggers a permissions review. They don’t change roles through HR workflows. They keep running with whatever permissions they were granted on day one until someone remembers to check.

The result is a growing population of privileged digital workers operating with far less oversight than we would extend to any human employee with equivalent access.

Agent risk concentrates around three recurring failure patterns. None require a sophisticated attacker. They emerge from ordinary operational pressure.

Shadow Agents: Agents deployed without formal review, proofs of concept that quietly become production infrastructure. The pattern mirrors shadow IT, but the stakes are higher because agents act, not just store data. OWASP has begun cataloging this risk class in its guidance on nonhuman identities, flagging unmanaged automated agents as one of the fastest-growing attack surfaces in enterprise environments.

Blank-Check Access: Over-provisioning happens because scoping permissions precisely takes more effort than granting a broad standing role. That’s how the database agent I described ended up with developer-level access for a task that needed almost none. Gartner projects that by 2027, a significant share of AI-related security incidents will trace to misuse enabled by overly broad access—not external compromise.

Moving-Target Agents: Agents are constantly reconfigured as teams iterate. An agent built to summarize support tickets gets expanded to draft responses, then send them, then issue refunds—each step reasonable in isolation but none individually triggering a permissions review. By the time anyone looks closely, the agent’s actual capabilities bear little resemblance to what was originally approved. Under the EU AI Act, organizations remain accountable for what their autonomous systems do, regardless of how incrementally those capabilities evolved.

Each failure shares a common root: a lack of enterprise visibility into what agents exist, what they can do and whether their access still reflects their purpose.

The Model Context Protocol (MCP) and similar standards make it dramatically easier for agents to connect to enterprise tools and data sources. That’s genuinely valuable. It also accelerates the risks above.

Lower integration friction means agents can be connected to new systems faster than governance processes can track. A protocol that makes it trivial to grant access also makes it trivial to grant that access without formal evaluation. MCP doesn’t create the governance gap; it widens it by removing the technical friction that used to slow unmanaged access down.

This is also where authentication and authorization blur dangerously. Authenticating an agent, confirming it is who it claims to be, is a solved problem. Authorizing what that agent should be allowed to do, and continuously verifying that authorization still matches its current task, is not. Most enterprises can confirm an agent connected successfully. Far fewer can confirm it should have or that its access today still reflects its purpose today.

​I didn’t arrive at this framework from a whiteboard. It came from watching the same incident repeat across enterprise environments: an agent with too much access, discovered too late, with no auditable history of what it actually did. The controls needed to prevent it are consistent.

Discover every agent. You cannot govern what you cannot see. Enterprises need continuous discovery that captures every active agent, including shadow deployments that never went through formal registration.

Assign unique identities and purposes. Every agent needs a distinct identity tied to a defined purpose. If you can’t answer “What is this agent for?” in one sentence, it shouldn’t have access to anything yet.

Trace the full chain of action. Enterprises need an auditable record of what each agent did, when and why, not just whether it authenticated. That chain of custody is what turns incident response from guesswork into an investigation.

Enforce policy at the point of access. Governance can’t live only in a policy document. It must be enforced in real time, at the moment an agent requests data or attempts an action, checked against its defined purpose before execution, not after.

Replace standing access with just-in-time, purpose-scoped permissions. Grant agents the minimum permissions required for a specific task, for the duration of that task and nothing more. Standing privilege is a liability that accrues risk every day it goes unused.

Accountability for autonomous system behavior is already here and growing. Organizations that get this right will know which agents are running, what they can do and whether that access still makes sense. That starts with an honest audit and governance built on defined roles, scoped permissions and continuous oversight.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

Original Source
https://www.forbes.com/councils/forbestechcouncil/2026/11/18/your-ai-agent-is-a-contractor-with-a-master-key-five-controls-enterprises-need/
Visit Forbes ↗
SHARE STORY:
𝕏 f in

Related Coverage in Technology