
Data Breaches Explained: How Unauthorized Access Turns Into Data Exposure
Sensitive data can be exposed through human identities, service accounts, applications, AI systems and network paths. Learn how those routes shape breach scope, investigation and response.

DATA BREACH DEFINED
A data breach is an incident in which protected or sensitive information is accessed, acquired, used or disclosed outside its authorized boundaries, whether through malicious activity, misuse or accidental exposure.
Sensitive data rarely sits behind a single access path anymore. Employees query it directly, applications reach it through service identities, APIs expose portions of it to other systems and AI workflows may retrieve it, pass it to a model and invoke downstream tools in the same run.
When one of those paths is compromised, the breach is defined by what happens to the data across that chain. What was actually queried, copied or disclosed, and by which identities? Did sensitive information stay inside the original system, or did it move into logs, downstream services or another environment?
Those answers shape the investigation, the containment work, whether compromised credentials need to be revoked or rotated, whether affected systems need to be isolated, whether customers or regulators may need to be notified, and how much business disruption the incident creates.
What is a data breach?
A data breach is a security incident in which sensitive, protected or confidential information is accessed, acquired or disclosed without authorization. Unauthorized use by someone who legitimately has access can qualify as well — having access to data doesn’t automatically make every use of it permissible.
The National Institute of Standards and Technology (NIST), for example, defines a breach broadly enough to include loss of control, compromise, unauthorized disclosure or unauthorized acquisition of personally identifiable information (PII) as well as an authorized user accessing PII for an unauthorized purpose.
Initial access is only the first part of the incident. The real impact depends on what the attacker can reach next, dictated by the permissions of the compromised account or service.
Data theft is only one possible outcome of unauthorized access. Incident response has to identify what the attacker actually did with the access they gained. Query history, object access, network telemetry and downstream logs help show how far that activity progressed and how confidently the organization can characterize the breach. From there, the team can scope the incident and determine what containment and notification work comes next.
It’s important to note that not every cyberattack involves a data breach. A distributed denial-of-service attack that takes a service offline is primarily an availability incident unless the event also involves unauthorized access to protected information.
See how Snowflake Trust Center helps surface security risks:

Data breach vs. data leak vs. security incident
A security incident is the broadest category. It covers events affecting the confidentiality, integrity or availability of information or systems, from credential compromise and malware to service disruption. The terms data leak and data breach both focus on exposure of protected information, but the line between them isn’t always consistent across regulatory frameworks or industry usage.
A data leak often refers to information being exposed unintentionally or through negligence — storage location accidentally configured for public access, an internal file shared with the wrong partner or sensitive data accidentally returned by an application.
A data breach generally involves unauthorized access, acquisition or disclosure of protected information. In some cases, a leak turns into a breach when an unauthorized party accesses the exposed data.
Regulatory definitions vary but are specific. GDPR, for instance, defines a personal data breach broadly enough to include accidental loss, alteration, destruction and unauthorized disclosure or access. HIPAA uses a different test: an impermissible use or disclosure of unsecured protected health information is generally presumed to be a breach unless the organization can show a low probability that the information was compromised. U.S. state breach-notification laws add their own definitions and triggers. Given these varying standards, each organization’s legal team should determine how each regulatory definition applies to its specific circumstances.
Common causes of data breaches
Data breaches begin through different paths, including compromised identities, exploited vulnerabilities, application-layer weaknesses and unintended exposure. In other cases, data is intercepted in transit or exposed through an application that already has legitimate access to it. Once access is established, attackers may collect or exfiltrate data, escalate privileges or use the initial foothold to reach additional systems.
Compromised human identities
Valid credentials are useful to attackers because, at least initially, the activity can resemble ordinary account use. Passwords obtained through phishing, credential stuffing or information-stealing malware can provide direct access to an employee account, while stolen session tokens may bypass the need to authenticate again.
If a compromised identity has access only to the objects and operations required for its role, the attacker inherits a relatively narrow set of effective permissions. An over-privileged account gives the attacker more to work with, including access the employee may no longer need.
Compromised service accounts and machine identities
Applications, workloads and automated services also authenticate to data and other systems, often without an interactive user behind them. Their credentials may include API keys, service-account passwords, tokens or workload identities, and the permissions attached to them can be substantial.
Long-lived secrets and broadly scoped service accounts are especially consequential when compromised because the attacker inherits the service’s access. Machine-generated activity may also occur at much higher volume than human activity, which changes what responders look for when distinguishing normal workload behavior from abuse.
Phishing and social engineering
Phishing attacks increasingly target more than passwords. Attackers might go after session tokens, convince users to approve fraudulent authentication requests or steal credentials for cloud and SaaS services.
That activity can be difficult to distinguish from legitimate use when it comes through a valid identity. This is where authentication and authorization start working together. Authentication policies govern how an identity signs in, while access policies determine what that identity is allowed to do once the session is initiated.
Vulnerabilities and application-layer exploitation
Software vulnerabilities provide another route to sensitive data, especially in internet-facing applications, APIs and services. According to Verizon’s 2026 Data Breach Investigation Report (DBIR), vulnerability exploitation accounted for 31% of breaches, up from 20% the previous year.
The compromised application may also already have a trusted connection to the data platform. In that case, an attacker does not necessarily need database credentials of their own; they may be able to use the application’s existing identity and permissions to reach the underlying data.
AI applications and agents
AI applications complicate breach scoping because sensitive data may move through retrieval systems, model context, tools and downstream services in a single workflow. Check Point’s 2026 Cloud Security Report found that 44% of organizations couldn’t trace where sensitive data went once it entered an AI workflow, while 61% reported either no visibility or uncertainty about lineage through AI processing.
Training and fine-tuning data, retrieval corpora, embeddings, prompts, retrieved context, application logs and agent tool calls can all contain or expose protected information. In enterprise systems, the AI application may also authenticate to data sources through service identities that have broader access than the end user.
RAG is one example. A user may never connect directly to a sensitive document store, but a retrieval service can surface documents through the application if retrieval-time authorization is too broad. Agentic systems extend that path further: an agent may query data, call APIs or invoke downstream services using the permissions attached to its tools and service identities. Prompt injection or a compromised tool chain can then turn those existing permissions into a route to data the attacker could not otherwise reach.
AI data security requires following activity beyond database access alone. Teams may need to examine retrieval logs, model inputs and outputs, agent tool calls, application telemetry and the identities used to access downstream systems to determine what data was exposed and how it moved through the AI workflow.
On-path interception
Data may also be exposed while moving between systems. If an attacker can position themselves on the communication path and transport protections are missing or incorrectly configured, traffic — including sensitive data or credentials — may be intercepted or altered.
Transport Layer Security (TLS) encrypts data in transit, while certificate validation helps ensure that clients are communicating with the intended endpoint. These controls protect the connection itself, rather than limiting what an authenticated identity can access after the connection is established.
Human error and accidental exposure
A breach does not require malicious intent. A data set shared with an unintended recipient, overly broad storage permissions or credentials committed to a public repository can all expose sensitive information without an attacker targeting an account.
What data could be exposed in a breach — and why the data type matters
Once investigators identify unauthorized activity, the type of information involved shapes what happens next. For example, access to a customer contact table creates a different response problem from access to API keys: one raises privacy and notification questions, while the other might give an attacker another route into the environment. Because statutory disclosure and regulatory obligations vary significantly by data type and jurisdiction, teams should consult legal counsel early to determine their specific response requirements.
Personally identifiable information
Customer and employee PII often includes names, addresses, identification numbers, contact information and financial account details. When PII is exposed, incident teams need to determine which individuals were affected, which fields were involved and which notification requirements apply. In the U.S., for example, state breach laws vary in the types of information they cover and the conditions that trigger notice.
Protected health information
Protected health information (PHI) brings healthcare-specific regulatory requirements, including obligations under HIPAA for covered entities and business associates in the United States. Security and legal teams need a defensible account of the affected records, the people involved and how the exposure occurred.
Credentials and secrets
Passwords, API tokens, cryptographic keys and service-account credentials deserve immediate attention because they play two roles in an incident: They are sensitive information themselves, and they may open the door to additional systems. Response teams therefore need to understand where each exposed secret is accepted, revoke or rotate it where appropriate and investigate any activity tied to those credentials.
Business data and intellectual property
Customer lists, source code, query results, pricing information and internal strategy documents may fall outside consumer breach-notification regimes while still representing serious business exposure. The response then centers on the sensitivity of the information, contractual obligations and the operational or competitive consequences of unauthorized disclosure.
What a data breach costs
IBM’s 2026 Cost of a Data Breach Report puts the global average cost of a breach at $4.99 million, up 12% from $4.44 million in its 2025 study. In the United States, the average reached $11.5 million.
That cost includes much more than a ransom payment or regulatory penalty. Investigation, escalation, notification, remediation, operational disruption and lost business all contribute.
Longer breach lifecycles also increase the scope of the response. As activity continues, teams may have more sessions, systems and data access to investigate, while containment, recovery and business disruption continue to consume resources.
Scope affects the workload of incident response directly. If one compromised account has broad access across the data environment, investigators have more objects, queries and possible records to trace. Narrower privileges reduce the number of legitimate paths available through the data.
How teams limit exposure after access is compromised
Different controls address different parts of the path to sensitive data. Authentication controls govern how human and machine identities connect, network policies restrict where connections can originate, encryption protects data in transit, and least-privilege access controls limit what an authenticated identity or application can reach. Logging and access history provide the evidence teams need afterward to determine what actually happened. Data security depends on all of these controls working together.
Strengthen authentication for human users
Multi-factor authentication (MFA) reduces reliance on a password as the sole proof of identity, while authentication policies define which methods users are allowed to use.
MFA should be enabled by default for newly created human users, alongside ongoing measures to phase out single-factor password authentication. Authentication policies give administrators additional control over the methods allowed for different users.
Manage machine identities and secrets separately
Machine identities are becoming a larger part of the exposure surface. The Check Point research also revealed that 48% of organizations named managing nonhuman identities as an AI-related identity challenge, while 40% cited API key and credential sprawl and 37% cited over-privileged service accounts.
Service accounts, API keys and workload identities have different authentication patterns from human users. They may operate continuously, authenticate noninteractively and require access to data or APIs that a person would use only occasionally.
Their controls should reflect those differences: short-lived credentials where possible, secure secret storage and rotation, narrowly scoped permissions and clear ownership of each machine identity. If a service credential is compromised, those limits determine the permissions the attacker inherits.
Restrict where sessions originate
Network policies restrict which network locations are allowed to connect to the environment. If an attacker tries to use valid credentials from a disallowed IP range or network location, the connection can be denied before authentication succeeds.
Used alongside authentication policies, network policies narrow the set of conditions under which a session is accepted in the first place.
Protect data with encryption and key controls
Encryption protects sensitive data when it’s stored and while it moves between systems. TLS covers data in transit, while encryption at rest protects stored data, backups and snapshots against direct exposure from underlying storage.
The protection also depends on how keys are managed. If an attacker can access both the encrypted data and the keys needed to decrypt it, encryption provides much less separation. Key-management controls need to govern where keys are stored, who or what can use them, and how they’re rotated.
For especially sensitive fields, organizations may also use tokenization or application-level encryption so that plaintext values aren’t broadly available even to systems that can access the surrounding record.
Limit privileges with role-based access control
Once a user or service is authenticated, role-based access control (RBAC) determines which objects and operations it’s authorized to use. Applied with least privilege, RBAC limits each identity to the access required for its role rather than granting broader permissions by default.
If a role only needs a few tables to complete assigned work, it shouldn’t have access to more. Keeping privileges narrow reduces the effective permissions an attacker inherits if that identity is compromised, and limits how much sensitive data is reachable through the account.
Those privileges also need maintenance. Roles tend to accumulate access as applications change, employees move between responsibilities and new data products are created. Periodic review helps keep the permissions an identity has aligned with the work it still needs to do.
Control what authorized sessions can see
Table-level access is only part of the picture. A person, service or application may need access to a table without needing every row or every value in plaintext. Row access policies restrict which records are visible under a given policy condition, while masking policies protect sensitive values within columns.
An analyst might need to query a customer table without seeing a full Social Security number. An application identity may likewise need only the fields required for a particular transaction. Keeping those boundaries close to the data limits what the consuming application can expose if it’s later compromised.
Masking lets the query run while withholding values the user doesn’t need to see. Row-level controls apply a similar idea to records, narrowing which portions of the table are visible to a given role or context.
Data loss prevention (DLP) adds another layer of protection by identifying sensitive data and applying policies related to how it should be stored, accessed and moved. Depending on the deployment, DLP can monitor or block risky handling of data at rest, in use or in transit.
Bound application and AI access to data
Applications and AI agents should receive only the data and tools required for the task they’re performing. That includes the permissions attached to the application identity as well as the context, retrieval sources and tool calls available to the AI system.
For agentic applications, tool permissions deserve particular scrutiny because the agent may be able to query data or invoke downstream services on a user’s behalf. Narrowing those permissions limits the actions available if the application is manipulated or its identity is compromised.
Preserve evidence for investigation
Preventive controls limit what an identity, application or service is allowed to do. Investigation depends on a different set of signals: authentication logs, network telemetry, query and object-access history, application logs, API activity and, for AI systems, retrieval traces and tool-call records.
Those sources help responders distinguish potential exposure from observed activity. A permissions map shows what an account could access; logs and traces help establish what it actually touched, whether data moved beyond the original system and how far the intrusion progressed.
Evidence also needs enough context to be useful. Time stamps, identity information, object names, policy decisions and downstream activity all help teams reconstruct the sequence of events and support a defensible breach scope.
Important Note: These practices don’t make breaches impossible. However, they do help teams narrow exposure up front and reconstruct activity afterward.
QUICK TIP
Review effective permissions, not just assigned roles. Inherited privileges and access accumulated over time can give a compromised identity more reach than its current function requires.
How Snowflake can support breach prevention and investigation
Within Snowflake, teams can govern how identities connect, narrow what they can access and trace activity after an incident. Authentication and network policies help govern how and from where identities connect, while RBAC, row access policies and masking policies limit what authenticated users and services can reach or see.
For investigation, Access History shows which objects queries accessed and which policies were applied, giving responders visibility beyond a static permissions map. Trust Center surfaces configuration and security findings that teams can review against organizational policies.
Follow the access path all the way to the data
Initial access explains how an intrusion started. Breach scope depends on what happened next: which identities and applications were involved, what permissions they carried, which data was reachable and what the available logs show about actual access or movement.
That is why breach readiness sits across identity, application, network and data controls. When those layers are well defined, responders have a much better chance of answering the questions that drive containment, notification and recovery without having to infer the scope from incomplete evidence.
KEY TAKEAWAY
The point of entry tells only part of the breach story. Scope depends on what the compromised identity, service or application could reach, what it actually touched and whether teams have enough evidence to reconstruct that activity.
Frequently Asked Questions
Your common questions about data security breaches, answered by Snowflake experts.
What should organizations investigate first after a suspected data breach?
Teams typically need to establish which identities, systems and applications were involved, what data they could access, what they actually touched and whether information moved outside the environment. That evidence helps determine containment, notification and recovery requirements.
Can AI applications cause data breaches?
Yes. AI applications can expose sensitive information through retrieval systems, prompts, model context, outputs, logs, agent tool calls or over-permissioned service identities. The model is only one part of the path; the surrounding application, data sources and tools also need to be included in breach scoping and investigation.
Explore Data Security Resources
Explore Data Security Topics
Deep dives into every aspect of data security

