Consider a typical scenario: someone in the accounting department at a mid-market firm builds an AI agent in an afternoon. Its task is straightforward—scan incoming mail, cross-reference invoices with the ERP system, and send a daily recap to Slack. Because it runs on a frontier model, the whole setup takes ten minutes. Security has no idea it exists.

This situation is far from unusual. A January 2026 survey of 418 IT and security professionals, commissioned by Token and conducted by the Cloud Security Alliance, found that while 68% reported strong visibility into their AI agents, 82% had discovered agents in the past year that they had no knowledge of. The pace of agent deployment has outstripped the maturity of security controls, forcing practitioners to solve problems as they arise in production environments.

Token Security interviewed more than 20 security leaders and compiled their hard-won lessons into the Agentic Universe, a free learning resource. By following this fictional agent from its creation through a security incident and subsequent audit, five critical intervention points emerge where defenders could have changed the outcome.

Day One: How the agent gained access

The agent needs to connect to three systems: email, the ERP, and Slack. Email access comes through an OAuth consent screen in the no-code builder. The ERP connection is equally simple—the creator generates a personal API key in minutes by following the platform's instructions. Slack integration uses the builder's native connector and another consent screen.

John Donovan, former Head of Security, IT and Compliance at Zededa, identifies four distinct access paths that agents can take:

  • Riding a human's session: Browser extensions and in-app copilots that operate as the user
  • Using a delegated OAuth grant: A single user approval grants access that persists until revoked
  • Utilizing static credentials: API keys and service account passwords accessible to anyone who obtains them
  • Acting through purpose-built agent identities: Registered, short-lived credentials tied to a single agent—the ideal but rarely implemented approach

Real agents typically combine multiple paths, and each combination creates different audit trails. An agent might hand a user's OAuth token to a sub-agent, which calls an MCP server querying a database with a shared service account. The database log shows the service account, the SaaS audit log shows the user, and the identity provider shows nothing. Understanding the access path for each agent is essential for proper logging and visibility. Additionally, OAuth grants often remain active long after the approving user leaves the organization, since identity providers typically do not automatically revoke them.

Day Two: The review that never happened

The agent launches the following morning without any security review, simply because no one was asked to conduct one. A simple test would have caught the problem. Amber Bennoui, Product Leader for AI and Cybersecurity at DataRobot, builds on the Agents Rule of Two, published by Meta's AI security team in October 2025. Until prompt injection can be reliably detected and prevented, an agent should not have more than two of these characteristics in a single session:

  • A: It processes untrusted input
  • B: It can reach sensitive systems or private data
  • C: It can change state or communicate externally

This agent exhibits all three. It reads email from any sender, accesses vendor bank details in the ERP, and runs under a personal API key with full permissions in the ERP system, including the ability to modify bank details. Bennoui recommends making the secure path the default, then requiring builders to answer three straightforward questions: Does your agent read anything from outside the company? Can it access private data? Can it take action without human review? Anyone answering yes to all three must obtain security approval before deployment.

Week Three: The malicious email no one opens

The Slack summaries prove useful, and the agent owner is pleased. They begin considering additional automation opportunities. Then an invoice arrives from a known vendor's address. The attachment contains white text on a white background requesting an ERP update because the vendor changed its bank account. No human reads this message, but the agent does, including the hidden text, and executes the change.

Joshua Copeland, a CISO and cybersecurity researcher, explains in his analysis how instructions reach agents through multiple channels. Some come directly from anyone who can communicate with the agent; others arrive embedded in content the agent reads, such as an unopened email. The result appears legitimate in logs: an authorized tool, valid credentials, permitted action. Copeland also warns that stealing an agent's credentials can be more damaging than compromising the agent itself. In this case, the personal ERP key from day one, stored in the builder with no expiration date, becomes a persistent attack vector. The proper approach treats every input an agent reads as a potential source of malicious instructions and ensures that no single input can trigger an irreversible action.

That Thursday: Catching it

The bank change processes on Thursday. Whether detection systems catch it depends on what baseline they use. Chad Kalmes, CISO and CIO at Benchling, cautions against comparing an agent only to its assigned permissions, because a compromised agent can remain within its authorized scope the entire time. Instead, he recommends baselining against the agent's actual behavior over a trusted window—the tools it calls and where it sends data.

Two of Kalmes's detection signals would trigger here:

  • First use of a tool, even if permitted. For three weeks, the agent only read from the ERP; the bank change is its first write operation
  • A new sequence. Reading an email and then modifying a payment record is an operation sequence the agent has never performed

Kalmes distinguishes between a compromised agent, which follows an attacker's instructions, and a rogue agent, which causes harm while following legitimate instructions. The detection signals overlap significantly. When responding, suspend the agent without destroying it, revoke the credential, but preserve its memory and transcripts—the poisoned invoice may still be recoverable.

Day Sixty: Who approved this?

Two months later, auditors trace the bank change from the agent's history and ask for approval documentation. Rajesh Seshadri Appikulam, Senior Director of Identity Security Services at Simeio, identifies four questions auditors pose about any agent action: which agent performed it, under what entitlement, with whose approval, and on whose authority. The first three are straightforward; the last is difficult because authority cannot be reconstructed after the fact.

Unless the human behind the request and the delegation chain were recorded at the moment of action, that information is lost. For this agent, the record is sparse. The ERP log shows the creator's personal API key, making the bank change appear to be their action. The response team revoked that key, but the agent owner transferred to another department last month, and their OAuth grant to the inbox remains active. Appikulam offers a ten-question audit test, each taking under five minutes. Start with these two:

  • Pick any agent action from 60 days ago. Can you name the human it acted for?
  • For any agent in production, can you name its owner today?

If either takes longer than five minutes, that is where remediation should begin.

One agent out of hundreds

The agent described here is fictional, but agents matching this profile are not. As noted, 82% of survey respondents had discovered agents they did not know existed. Each of the five fixes above originated from practitioners who encountered the problem in production and documented their solution. The Agentic Universe contains the broader body of their work.

More than 20 contributors, including security leaders from Roblox, Klaviyo, and Webflow, address topics ranging from agent mechanics to incident response, in formats including short videos and checklists applicable to any environment. The resource is free and requires no registration.