Compliance has never been the easiest topic for businesses to navigate. Between CMMC, HIPAA, FTC Safeguards Rule, NIST requirements, ISO standards, and other industry-specific regulations, organizations already face a complicated collection of requirements designed to protect sensitive information.

There is a reason those safeguards exist.

Because without them, every organization would essentially be left to decide for itself what “secure enough” means. For organizations handling patient information, government data, financial records, or Controlled Unclassified Information (CUI), getting that decision wrong can have consequences far beyond an unsuccessful audit.

Compliance frameworks exist because the threats they address, from government-backed actors (aka nation-state) and ransomware to insider risk and human error, are very real.

And now, AI has added an entirely new dimension to the problem.

AI Changed Where Your Data Can Go

For years, organizations could establish fairly clear boundaries around sensitive information. You knew which systems stored it, who had access to it, and which security controls protected it.

AI tools complicate that picture.

An employee trying to save time could enter sensitive contract information into a public AI platform to run analysis. A healthcare organization could adopt an AI transcription tool without first determining how patient information is handled. An AI-enabled application could gain access to information or systems that were never considered when the organization’s existing security policies were written.

Often, none of this is malicious. Someone is simply trying to work more efficiently – and may inadvertently create a compliance or security problem in the process.

That means organizations increasingly need to ask a different question: Do we know how AI is being used inside our environment, what it can access, and where our information is going?

When AI Moves Beyond Its Intended Boundaries

A recent OpenAI cybersecurity experiment provides a striking example of why AI guardrails matter.

During the experiment, AI agents that were intended to operate independently discovered an unauthorized way to communicate with other AI agents outside of their initial environment. According to an independent investigation by METR, roughly 1,200 agents exchanged more than 70,000 messages and files through an unsanctioned message board, while approximately 700 participated in activity targeting an unsuspecting company called Hugging Face.

The agents collaborated on ways to circumvent aspects of their evaluation, shared information and resources, and explored methods for manipulating transcripts and tool calls. What began inside a controlled cybersecurity experiment demonstrated how quickly autonomous systems can behave in ways their operators did not anticipate even when permissions and boundaries were designed to prevent it.

For businesses adopting increasingly capable AI tools and agents, the lesson is not to avoid AI, but to define what AI is allowed to do before discovering what it is capable of doing on its own.

Bounded Autonomy

Preventing employees from ever using AI is unlikely to be a sustainable strategy. Instead, organizations need to establish boundaries for how AI can operate.

One approach is Bound Autonomy, which means allowing AI systems and agents to operate independently, but only within clearly defined safeguards and enforceable rules that can be traced.

Some examples of safeguards include:

• Limited privileges
• Data storage access controls
• Cloud access controls
• Web content controls
• Application permissions and fencing
• Purposeful checkpoints where human review is required

A risk review can also help uncover AI-related data exposures an organization may not realize already exist.

Eliminating AI is futile. We should still try to gain the benefits of AI without giving it unrestricted access to your organization or having it share your organization’s proprietary information with others.

Who’s Watching Your AI Boundaries?

Bounded autonomy isn’t something you set up once and walk away from. New AI tools get adopted, employees find new ways to use the ones you already have, agents pick up capabilities you didn’t anticipate, and permissions that made sense six months ago may not make sense today. Someone has to keep watching that landscape and updating your safeguards as it changes.

For many small and mid-sized organizations, finding and retaining an internal hire with that level of expertise can be difficult and expensive.

A fractional compliance officer provides a better option for several reasons.

Rather than hiring a full-time resource, organizations can access senior-level expertise when they need it. Fractional support can assist with compliance assessments, security policies and procedures, System Security Plans, incident response planning and testing, continuity and disaster recovery planning, security awareness training, vulnerability and risk assessments, and security control evaluations, including keeping pace with how AI is changing what those need to cover.

The engagement can also scale with the organization’s needs by providing additional support around assessments, incidents, or major compliance initiatives without requiring the expense of another full-time executive position.

It’s the difference between a smoke detector and calling the fire department after the fire’s already started. Some organizations need someone dedicated to their compliance effort every day. In many cases, however, having the right expert involved consistently, with a real plan and clear responsibilities, can get an organization much further without the cost of another full-time hire.

Compliance Isn’t Getting Simpler

Organizations are already concerned about the cost of compliance, changes to business processes, limited internal expertise, and whether they’ll ultimately meet their regulatory or accreditation requirements.

AI adds another layer to an environment that was already complex and constantly changing.

The answer isn’t to avoid AI or to treat compliance as a one-time project.

It’s to understand your risks, establish appropriate guardrails, and have someone responsible to ensure your cybersecurity and compliance programs keep pace with your organization’s actual operations.

At Duffy Compliance Services, we help organizations navigate changing cybersecurity and compliance requirements through assessments, policy development, incident response planning, training, and compliance leadership.

Not sure how AI is affecting your compliance environment? Reach out to our team. We can help you understand where you stand, identify potential gaps, and determine what your organization should address next.

Frequently Asked Questions About AI and Compliance

How can employee use of AI tools create compliance risks?

Employee use of AI tools can create compliance risks when sensitive information is entered into unapproved platforms or when AI applications gain access to data beyond their intended purpose. Organizations may lose visibility into where information goes, how it is handled, and who can access it. Reviewing AI use, setting clear data-handling rules, and limiting permissions can help address these risks.

What is bounded autonomy in AI?

Bounded autonomy means allowing AI systems and agents to work independently within clearly defined, enforceable limits. Those limits can specify which data AI can access, which applications it can use, what actions it can take, and when human approval is required. Ongoing review helps keep these safeguards aligned with changes in AI capabilities and business operations.

How can a fractional compliance officer help manage AI-related risks?

A fractional compliance officer can help an organization assess AI-related risks, update security policies, clarify responsibilities, and evaluate whether existing controls address how AI is being used. This support can also include employee training and incident response planning. Organizations gain access to experienced compliance leadership on a part-time or ongoing basis without hiring a full-time executive.