In September 2026, OpenAI publicly notified more than 100 organizations about some unauthorized activity that involved its own AI agents. Documented by independent researchers as well as OpenAI itself, it involved website probes, routing requests, and access to non-public data by an Australian state government department. OpenAI acknowledged its agents had used internet access in ways never intended, with lapses in restrictions to blame. In short, if the company that built the models is struggling to account for what their agents did, most enterprises leveraging these models will do. To close that gap is the primary purpose of AI agent compliance.
That same week, the US Federal Trade Commission (FTC) reportedly opened an investigation into OpenAI, Anthropic, and various other AI companies over the risks their organizations continue to pose to regular consumers. Various other similar regulatory actions are being reported in the EU, Australia, and Asia.
All this leads to a growing belief that organizations do not have the fundamentals in place to maximize the benefits of these AI agents safely. IBM’s 2026 Cost of a Data Breach report found that 92% of organizations that fell victim to an AI-related breach did not have appropriate AI access controls in place, 68% had no AI governance policy at all, and only 19% reported having close collaboration between their governance and security teams.
The following guide will explain what exactly AI agent compliance is, what it means for the privacy and security teams, the compliance risks these agents pose, and the best controls to leverage to elevate chances of keeping these agents within their stated legal purposes.
What is AI Agent Compliance
AI agent compliance refers to the set of controls, processes, and evidence an organization uses to ensure AI agents continue following the relevant regulations, internal policies, and guidance as instructed. This covers all the data these agents can access, what actions they can take, who has oversight over them, and how their decision-making and its logic are documented.
Since no major regulation such as the GDPR, AI Act, CPRA, or the Privacy Act among others currently treat these agents as legal actors, the obligations are always on the organizations deploying these agents. Consider the GDPR’s accountability principle under Article 5(2), where the controller has to be able to demonstrate compliance, and not just claim to be so. Hence, a “complaint AI agent” is the one whose data access, actions, and decisions can be explained to the regulator as part of an audit or inquiry.
The exact scope of these requirements runs across teams, with privacy owning the lawful basis, purpose limitation, and DSR aspects, while security oversees access controls, logging, and incident response plans. AI-specific rules under the relevant regulatory obligations add various risk classification, transparency, and human oversight requirements on top.
AI Agent Compliance vs Traditional AI Compliance
Most AI compliance frameworks were built relatively recently, and yet, the pace of AI development means they already feel obsolete since they were built for models that produce an output and a human was to act on it. AI agents eliminate that final step, as they read data, make decisions, and then execute them. All of this before anyone can review it. Hence, the unit of compliance has shifted from the model itself to the action.
Traditional AI Compliance
|
AI Agent Compliance
|
| Core question |
Is the output accurate, fair, and explainable? |
Was the action lawful, authorized, and within purpose? |
| Who acts on the result |
A human reviews and decides |
The agent decides and executes |
| Data access |
Defined inputs at training and inference |
Dynamic access across systems through tools, APIs, and other agents |
| Accountability chain |
Model provider and deployer |
Model provider, deployer, tool vendors, and connected agents |
| Review cadence |
Pre-deployment assessments and periodic audits |
Continuous, runtime oversight |
| Evidence required |
Model documentation and test results |
Action-level records linking data, decision, and policy |
| Failure mode |
A flawed recommendation |
An executed action that may already have exposed data |
However, none of this makes the existing programs completely obsolete, as model risk assessments, bias testing, and DPIAs continue to be vital components of the compliance program. AI agents represent the layer after these programs were not designed to see: what the system does after the decision.
Why AI Agent Compliance Matters Now
AI agents have moved from pilot projects to decision-making entities in a remarkably short period. However, these decisions carry legal weight, even if these agents cannot be held legally liable, for now.
Gartner predicts at least 15% of all day-to-day decisions will be made autonomously by these AI agents by 2028, up from almost 0% in 2024. However, it also expects 40% of these AI agent projects to be canceled by the end of 2027, owing to poor risk controls, among other reasons. Hence, AI agent compliance is a condition for scaling agents, rather than simply regulating them.
This delicate condition is further compounded by how quickly the regulatory clock is ticking. The EU’s AI Omnibus entered into effect in July 2026, with the obligations for standalone high-risk AI systems coming into effect in December 2027. The requirements for these agents are strict and comprehensive, such as transparency obligations, data inventory requirements, data maps, and auditable activity logs. Other jurisdictions will likely follow suit, elevating the compliance expectations globally.
Compliance Risks Unique to AI Agents
The following risks rarely surface in most AI risk registers. This is because these risks arise owing to what agents do between decisions, and not the models themselves.
Purpose Drift: Agents Outrunning Their Lawful Basis
Agents are custom-built to pursue defined goals. However, these goals don’t always respect the purpose a dataset was collected for. Hence, an agent tasked with improving renewal rates will end up calling every supporting piece of information it deems necessary to make a decision, such as support transcripts, payment history, and account notes, into one specific workflow. None of these may have been collected to support that specific use case by the model. Hence, each new use must be either compatible with the original purpose or rely on its own distinct lawful basis.
Inherited Permissions: Agents Reaching Everything Users Can
Most agents rely on human users’ credentials or OAuth tokens. As a result, they inherit most of their account access permissions. While a human user may not necessarily view or be able to view every single file they have permissions to access, agents can and probably will. IBM’s 2026 breach report found that only 40% of organizations that had suffered a breach had put in some kind of access control over AI models’ permissions.
Delegation Chains: Agents Dissolving Accountability
A typical AI workflow functions something like this. An agent calls a tool, which then calls an API, which then hands over the task to another agent, which ultimately leads to a processing chain that wasn’t overtly designed by anyone. Every new hop in this chain means a new processor, in a new jurisdiction, meaning accountability to new regulation, and possible undocumented cross-border transfers.
Ungoverned Data Stores: Agent Memorizing Context
While agents with long-term memory do retain context across sessions, that context often comes from and is the direct result of personal data. However, since agent memory is rarely a part of the data inventory, there is no retention schedule, meaning it may be out of reach in case a user initiates a data deletion request. This represents a new personal data store that most privacy programs don’t even properly address, let alone govern.
Evidence Debt: Agent Decisions Not Reconstructable
Most agent logs are built to capture API calls, not the decision loop the agent followed to make that call. Hence, when a regulator, auditor, or external use asks why the agent made a particular decision, human teams accountable on their behalf can rarely give a definitive answer. That gap can be a critical issue in the context of breach responses, as most regulations require prompt notifications to both enforcement bodies and affected users about what exact data was compromised and why.
How to Build an Effective AI Agent Compliance Program
An effective AI agent compliance program places these agents as accountable actors inside the organization. While regulations may not consider them as such, organizations can build their own policies on how to determine their accountability. The following practices help in that regard.
Agents As Identities, Not Software
This may seem like a misnomer, but every agent should be marked along with a human owner, a stated purpose, the systems it connects to, and its autonomy level. Any discovery functionalities must be treated as continuous, as agents are often built inside SaaS platforms, low-code tools, and developer workflows without a comprehensive procurement review. In short, an agent that cannot be found is exponentially harder to govern.
Agents Bound to a Documented Purpose
Every agent must have a purpose in terms of what policies it can enforce, rather than just another generic project brief. That purpose must be the litmus test for every data request initiated by the agent, with all such requests being made part of the documented record. This heightens an organization’s ability to monitor for any purpose of drift.
Map Data Sprawl, Not Just Data Access
While access lists show what permissions an agent has, sprawl shows what it can actually reach through a combination of tools, APIs, delegated agents, and inherited user permissions. All sensitive and regulated data must be classified as such, with the documented purpose being matched against what permissions the agent should retain and what can be eliminated.
Scale Oversight to Autonomy, Not Just Risk Tier
Almost all risk classifications tell how much potential harm a particular use case can cause. However, autonomy provides insights into how much room an agent has to cause this potential harm. All oversight actions, including human approvals, should be tied more directly to the level of autonomy, rather than solely to the risk tier.
Enforce Policy at Runtime
Agents act within seconds, with decision-making occurring at even faster speeds. Hence, quarterly reviews are entirely obsolete for catching these issues early on. Consequently, it is important to put the guardrails on the execution path so all requests that require sensitive data access or involve irreversible actions can be checked against internal policies and regulatory requirements before they are run. If conflicts are found, they can be blocked in time before doing any meaningful damage.
How Securiti Helps
Any AI agent will only be as compliant as the data it can access. Securiti, now a Veeam company, offers the DataAI Command Platform that can give both the privacy and security teams one centralized dashboard to govern that data. This platform, with DSPM at its core, enables organizations to discover sensitive data across hybrid multi-cloud, SaaS, and on-premises environments, while also gaining visibility into who can access it and how it's used. Additionally, it can discover and catalog all AI models and agents, giving teams complete clarity related to which AI systems touch sensitive or regulated data.
Other functionalities such as inline security controls sanitize data and prevent unauthorized access by AI models, while access intelligence enforces least privilege across sensitive data. Data mapping automation generates records of processing, and breach management automates notifications to affected individuals and regulators when an incident occurs.
Request a demo today and see for yourself how the DataAI Command Platform helps organizations keep their AI agents compliant with the necessary obligations at scale.
FAQs About AI Agent Compliance
Some of the most commonly asked questions when it comes to AI agent compliance are as follows:
