Skip to content
CARE

AGENTIC CARE · WHITEPAPER

The Agent That Followed Every Rule

Permitted steps can still form an unacceptable process

Every tool call was allowed. The combined report still put personal information where a wider audience could read it. The risk lived in the path, not in a broken permission.

All the checks can be green

Conceptual illustration of why step-by-step control misses the issue
Individually permitted steps can converge on an output the organization did not intend to publish.

An agent is authorized to read customer data, summarize trends, prepare an internal report, and store it in an approved workspace. No access violation is recorded. The workspace is broadly readable. The report makes personal information available to people who were never the audience for the source records.

This is an illustrative path, not a reported customer incident. It is the failure mode to assess: the process, not a sequence of isolated calls.

Permissions describe actions. Business exposure is often produced by the outputs, the consumers, and the next action those consumers take.

Look at the path, not the calls

CARE Process value-stream view for Accounts payable intake with Sources, Intake, Process steps, Outputs, and Consumers columns, plus Must and Never conditions.
Observed CARE value stream for accounts-payable intake: sources, steps, outputs, and named consumers on one view. The capture shows a path that can be reviewed as a path.

The review that catches this asks who can open the final output, whether personal data is still identifiable, whether the storage location inherits a wider permission, and whether a person is required before the output moves again. A demonstrated payables stream makes the same structure discussable: sources, steps, outputs, consumers, and the conditions that must hold. The structure is what you copy. The payables content is not your incident.

The owner changes the arrangement or accepts the exposure

The business owner of the report — not the owner of each tool — decides. Removing the tool is one option. It is often the wrong one.

  1. Draw the path for one agent that assembles information and stores or sends it: sources, steps, output, who consumes it.
  2. Mark the step where sensitivity increases. In the example, that step is the combined report in a shared place, not the original read.
  3. Choose a change: a narrower audience, redaction, a restricted location, or a person before external sharing. Or record that the residual exposure is accepted, by name, with the reason.
  4. If the change is a rule on paper only, label it as not yet a barrier. Assign the person who can implement it in the system that stores or sends the output.
Do not close the review with “the agent complied.” Close it with an accepted path or a recorded exposure.

Obedience is not the same as an acceptable process

You keep the useful agent. You stop treating a clean permission log as evidence that the operating arrangement is one you meant.

The result is one path diagram and one decision. The same review applies the next time a redesign changes a consumer without adding a tool — a common way a safe sequence becomes an unsafe one. Management attention moves to the arrangement, which is where the authority to change it already sits.

Take action now

  1. Tell your agent to register itself in CARE: https://ai-managementsystem.com/under-care.
  2. Let the agent establish the management baseline.
  3. Use the outcomes.

Images and text generated with the assistance of ChatGPT by OpenAI.

Download