When AI Can See Too Much
AI copilots, agents, and large language models can search company information and turn it into answers or actions. That convenience can make existing data-governance weaknesses more consequential: information that was technically accessible but difficult to find may become easier to locate through a natural-language prompt. The risk is not that every AI system automatically leaks data, but that AI can amplify weaknesses in access, identity, and data quality.
Company Data Breach Examples: From Oversharing to AI Exposure
The following hypothetical scenarios illustrate exposure paths; they are not reports of actual incidents and do not establish that a breach occurred.
- Overshared repository: A shared drive contains sensitive customer, salary, research, or source-code material with permissions broader than necessary. An AI assistant connected to the drive could surface that material to a user whose access is technically allowed but not justified by their work.
- Unapproved AI service: An employee submits confidential material to an app that the organization has not approved. This creates a potential data-handling concern that the organization would need to investigate, including what information was submitted and the service’s applicable settings and terms.
- Compromised account: Someone obtains a valid account and uses its permissions to reach company information. If those permissions are excessive, the account may provide access beyond what its owner needs; multifactor authentication and appropriately limited access help address this risk.
These examples are exposure paths, not evidence of unauthorized access or disclosure. Establishing whether an incident occurred requires evidence about what was accessed or disclosed, by whom, and when.
What the 1,000-Environment Assessment Found
A June 2025 BleepingComputer report summarizing Varonis’s sponsored State of Data Security Report described an assessment of data from 1,000 real-world IT environments. BleepingComputer reported that 99% of the assessed environments had sensitive data unnecessarily exposed to AI tools; 90% of sensitive cloud data, including AI training data, was open and accessible to AI tools; 98% had unverified apps, including shadow AI; one in seven did not enforce MFA across SaaS and multicloud environments; and 88% had ghost users.
These figures describe the assessed environments and the exposure or risk factors reported there. They do not represent all companies or prove that exposed data was accessed, disclosed, or involved in a breach. The assessment is a reason to review permissions, identities, sensitive repositories, and connected applications—not evidence of an incident at any particular organization.
Why AI Can Increase the Consequences of Old Weaknesses
AI can change discoverability. A person may receive a concise answer assembled from documents they could access but would not ordinarily find or read. This makes it important to review whether permissions are appropriate before relying on an assistant’s configured access rules.
AI can also change scale: an assistant or agent may search across connected sources or invoke tools as part of a workflow. Each connection and delegated identity should therefore have access limited to what its task requires. Separately, the integrity of AI outputs depends in part on the quality and reliability of the information supplied to the system. These are distinct concerns: human-to-machine access is what a person can retrieve through an AI tool, while machine-to-machine integrity concerns the information and actions passed between connected systems. Neither concern, by itself, establishes that a breach has occurred.
Practical Controls for AI and Data Security
Inventory AI tools, agents, model connections, and data sources, including approved deployments and unapproved applications identified through the organization’s review processes. For each integration, record its owner, the data it can reach, its permissions, and whether it can act on data or only retrieve it.
Classify sensitive information and review effective access. Remove broad links and inherited permissions without a current business justification. Apply least privilege to users, service accounts, connectors, and agents; remove unused accounts and credentials; and enforce MFA, particularly for accounts with administrative, cloud, or sensitive-data access.
Monitor data access and AI activity in context. Keep logs that help investigators determine which identity accessed which source, when, and through what application. A response plan should distinguish a risky configuration from evidence of unauthorized access and preserve relevant records during an investigation.
For AI systems connected to data or tools, assess the sources they use, protect configurations and credentials, limit available actions to the task, and require human review where errors could cause material harm. These measures complement access controls.
Use a Threat-Assessment Cycle, Not a One-Time AI Review
Treat AI risk review as an ongoing process. Before deployment, document intended use and affected parties, assess risks, and test controls. After launch, monitor the system and revisit the assessment when its models, data, permissions, or workflows change.
Prioritize reviews by sensitivity, reach, and potential consequence. Ask what the tool can access, whether that access is necessary, who can use it, and what could happen if an identity or connection were compromised. Test representative roles and prompts, record remediation owners and deadlines, and retest after changes. Testing can inform the review but cannot establish that every possible exposure has been eliminated.
Automation can help monitor data and remediate issues, but changes to permissions can have significant effects. Stage high-impact changes, review them, and retain a way to reverse them if needed.
Treat Exposure as a Signal to Act
The assessment’s reported findings point to conditions worth reviewing, not a conclusion that every organization has experienced a breach. Organizations can use them to examine data access, identity controls, AI applications, and connected systems. Clear ownership and continuing review help address unnecessary exposure while supporting the responsible use of AI.
Sources
ai can see too much