SECURING THE NEXT 250 STARTS WITH WHAT YOUR AI CAN TOUCH

Written By: Chet Hayes, Vertosoft CTO

The Next 250 Will Run on Data, and Increasingly on Agents

This year’s Cybersecurity Awareness Month theme from CISA, “Securing the Next 250,” asks a forward-looking question. As the nation marks its 250th anniversary, how do we secure the systems the next era will depend on? CISA’s answer centers on critical infrastructure and the state, local, tribal, and territorial governments that keep water, power, health care, and emergency services running.

Those systems are changing faster than most security programs. Agencies are no longer just piloting AI. They are connecting assistants and agents to case files, permitting systems, benefits data, and operational records. The next 250 years of public service will run on that data, and more and more of it will be read, moved, and acted on by software that no one is watching in real time.

Last October, in “Beyond the Calendar,” we argued that a secure foundation is what earns an agency the right to use AI safely. This year the question has moved on. AI is already inside the perimeter. The work now is securing what it can touch.

CISA gives critical infrastructure owners a simple framework for this year: the 3Rs of Reduce, Replace, and Recover. Read with AI in mind, those three words make a practical playbook.

From Answering to Acting

The first wave of generative AI answered questions. The current wave takes action. An agent can query a database, open a file share, draft and send a message, update a record, or call another tool, often with its own credentials and standing access.

That changes where risk lives. A chatbot that gives a wrong answer is an accuracy problem. An agent that can reach a folder of controlled unclassified information, or that follows instructions hidden in a document it was asked to summarize, is a security problem. Prompt injection is, in effect, phishing aimed at software, and agents do not hesitate before they click.

Attackers are using the same technology. AI shortens the time between gaining a foothold and finding something worth stealing. It helps an intruder read what they have found, decide what to probe next, and keep moving with less human effort.

The common thread is data. Whether the actor is an employee, an adversary, or an agent acting on someone’s behalf, the question that matters is the same: what sensitive data can it reach, and what can it do with that data? That is where the 3Rs come in.

Reduce: Shrink What AI Can Reach

CISA’s first R is reducing the attack surface. For most agencies, the largest unmanaged attack surface is no longer a server or a port. It is data: copies of sensitive records scattered across cloud storage, SaaS applications, shared drives, and old project folders, much of it over-permissioned and long forgotten.

AI makes that sprawl more dangerous. An agent connected to a broadly shared drive inherits every permission problem that drive already had. It can surface a document no human would have stumbled on, and it can do so in seconds.

Reducing this surface starts with three questions that many agencies still cannot answer with confidence:

  • Where is our sensitive data? Not where policy says it should be, but where it actually sits today, including the copies.
  • Who and what can reach it? That means people, service accounts, and every AI assistant or agent connected to the environment.
  • What is actually being used? Data and access that no one needs is pure risk. Removing it is the cheapest security control available.

Modern data security tooling now discovers and classifies sensitive data across environments automatically, and maps it against the identities, including AI agents, that can access it. The principle is old, least privilege, but it now applies to non-human actors that multiply faster than any team can review by hand. An agency that cannot inventory its agents and what they can touch has not yet reduced its attack surface. It has only stopped looking at part of it.

Replace: Retire Point-in-Time Governance

CISA’s second R is replacing end-of-support devices, the aging edge equipment that no longer receives patches. There is a governance equivalent, and most agencies are still running it.

The traditional model of data and AI governance is periodic. A committee reviews a use case, a policy document is approved, and an assessment is filed. That worked when systems changed on a quarterly release cycle. It does not work when an agent can be built in an afternoon, connected to new data the next morning, and take thousands of actions before the next review meeting.

Point-in-time governance is the end-of-support device of the AI era. It still looks like a control, but it has stopped protecting anything.

The replacement is governance that runs continuously and enforces at runtime:

  • Policies expressed as rules a system can check, not just documents people sign.
  • Ongoing assessment against the frameworks agencies already answer to, such as the NIST AI Risk Management Framework.
  • Enforcement at the moment of action, so a non-compliant agent request is stopped before it executes, not discovered in next quarter’s audit.
  • Automation of routine stewardship, so scarce governance staff spend their time on judgment calls rather than paperwork.

This mirrors the shift from a point-in-time Authority to Operate to continuous ATO that we discussed last year. The logic is the same. If the system changes continuously, the oversight has to as well.

Recover: Know What the Agent Touched

CISA’s third R is recovering quickly to sustain operations. Its broader guidance for organizations backs this up: turn on logging, back up and encrypt data, report incidents, and maintain an incident response plan that is actually exercised.

AI incidents add a question most response plans were not written to answer. When an agent misbehaves, whether through a bad configuration, a manipulated prompt, or a compromised credential, the first thing leaders will ask is: what did it touch? Which records did it read? What did it copy, change, or send, and to whom?

If an agency cannot answer that within hours, it cannot scope the incident, notify the right people, or restore trust in the system. Recovery depends on three capabilities:

  • The ability to trace where data came from, how it changed, and where it went, including through AI pipelines.
  • Activity records for non-human actors. Agent actions logged with the same rigor as privileged human users.
  • A rehearsed playbook. An AI-specific scenario in the annual tabletop exercise, with clear authority to suspend an agent’s access quickly.

Recovery is also about continuity. CISA asks organizations to plan for running mission-essential functions without critical systems. As agencies fold AI into frontline workflows, that plan needs a line for what happens when the agent is switched off.

Security and Governance Are One Loop

In many agencies, data security and data governance still sit in different offices, with different tools and different budgets. Security teams worry about exposure. Governance teams worry about quality, policy, and compliance. AI erases that line.

The 3Rs only work as a loop. Discovery tells governance what data exists and how sensitive it is. Governance turns that context into policy for what each AI system may use. Enforcement applies that policy at runtime. Lineage and activity records feed back into discovery and make recovery possible. Break any link and the loop fails: governance without discovery is guesswork, and discovery without governance is a list of problems no one owns.

The market is moving in the same direction. Data security platforms such as Cyera are extending from sensitive-data discovery into controlling what AI agents can access. Governance platforms such as Collibra are moving from catalogs and policy documents to runtime oversight of AI. The two even integrate, so classification from one enriches governance in the other. The specific tools matter less than the architecture. Agencies should expect their security and governance layers to share context, not sit side by side.

Your First Action for November

Last year we suggested starting with MFA. This year’s first action is just as concrete: inventory every AI assistant and agent in your environment, and write down what data each one can reach.

Include the obvious ones and the quiet ones: the copilot built into your productivity suite, the chatbot a program office stood up, the integration a contractor connected last spring. For each, note the data it can access, the identity it runs under, and who would know if it did something unexpected.

That single list touches all three Rs. It shows where to reduce access, which oversight needs replacing, and what you would need to trace if something went wrong. Expect to find agents no one tracked and access no one meant to grant.

Securing the next 250 years will not depend on any one technology. It will depend on knowing what our systems can touch and keeping that answer current. For AI, that work starts now.

Vertosoft works with public sector agencies and partners to bring data security, governance, and AI technologies to the mission through the contract vehicles they already use. To talk through where your agency stands, contact us at sales@vertosoft.com.