Artificial intelligence is transforming how insurance companies manage claims, evaluate risks, process policy applications, detect potential fraud, and interact with policyholders. From AI-powered document processing to intelligent claims assistants, insurers are increasingly integrating AI into critical business operations.
These systems can analyze accident photographs, extract information from medical reports, summarize policy documents, identify unusual claims patterns, and support underwriting decisions. More advanced AI agents can also interact with claims management platforms, customer relationship management (CRM) systems, payment applications, internal databases, and third-party services. However, greater AI adoption also introduces new security challenges.
Insurance companies manage some of the most sensitive categories of personal information, including policyholder identities, financial records, medical histories, insurance claims, accident reports, and beneficiary details. When AI applications and autonomous agents gain access to these resources, excessive permissions, insecure integrations, and inadequate governance can create additional exposure paths.
According to the National Association of Insurance Commissioners (NAIC), 92% of surveyed health insurers, 88% of surveyed auto insurers, 70% of surveyed homeowners insurers, and 58% of surveyed life insurance companies reported that they use, plan to use, or plan to explore AI and machine learning in their operations.
These figures demonstrate the growing importance of securing AI across insurance operations.
A real-world cybersecurity incident further illustrates the sensitivity of this environment. In June 2025, Aflac disclosed unauthorized access to its U.S. network. Its subsequent investigation identified personal information associated with approximately 22.65 million individuals in potentially impacted files, including claims information, health information, and other personal details. Aflac reported that social engineering was used to gain network access.
Although this incident was not reported as an AI-related breach, it demonstrates the potential consequences of unauthorized access to sensitive insurance information. As insurers introduce more AI agents and connected applications, securing the identities, permissions, and data accessible through those systems becomes increasingly important.
For insurance companies, AI security must extend beyond protecting individual models. Organizations need visibility into AI applications, sensitive data exposure, agent identities, third-party integrations, and the trust relationships connecting enterprise systems.
How Insurance Companies are Using AI
AI adoption spans multiple insurance operations, helping companies improve processing efficiency and support business decisions.
Common applications include:
- Claims processing: Extracting information from claim forms, analyzing accident photographs, reviewing supporting documents, and assisting with claim assessments.
- Underwriting: Analyzing applicant information, evaluating risk factors, and supporting underwriting classifications.
- Fraud detection: Identifying unusual transaction patterns, inconsistent information, and potentially suspicious claims.
- Customer service: Answering policy-related questions, summarizing coverage information, and assisting customers with routine requests.
- Policy administration: Retrieving policy details, processing documentation, and supporting policy servicing.
- Risk assessment: Analyzing historical information and other datasets to support risk evaluation and pricing.
The NAIC also reports that insurers use AI for accident image analysis, claims settlement estimates, fraud detection, prior authorization, and underwriting-related activities.
As these applications become interconnected, AI agents may gain access to multiple internal and external resources. This creates a need for stronger identity management, least-privilege access, data protection, integration security, and AI governance.
Major AI Security Risks in Insurance Companies
1. Policyholder Data Exposure Through AI Applications
Insurance companies collect and process extensive amounts of sensitive information. Depending on the insurance product, this may include medical records, financial statements, personal identification details, accident reports, beneficiary information, and claim histories.
AI assistants and agents may access these records to prepare claim summaries, retrieve policy information, analyze medical documents, or respond to customer queries. However, when these systems have excessive permissions or access to interconnected data repositories, sensitive policyholder information may become exposed beyond its intended business requirements.
Real-world scenario:
Consider a health insurance company using an AI assistant connected to its internal accounts and claims systems. The assistant retrieves hospital bills, diagnostic reports, and policyholder records to generate claim summaries and respond to customer queries.
However, the application identity also has access to medical records belonging to other customers and unrelated claims folders. If the assistant is manipulated or compromised, it could retrieve information beyond the intended policyholder's records. For example, a claim summary generated for one policyholder could unintentionally include another customer's medical history, personal details, or claim information in its response.
The security issue is not limited to the AI model itself. It also involves the identity behind the application, the sensitive data it can access, and the relationships connecting it to enterprise resources.
Grafyn's Sensitive Data Disclosure helps identify sensitive policyholder information that AI agents and connected workflows can access, highlighting potential data exposure risks across claims, underwriting, and customer service systems.
2. AI Agent Identity and Excessive Permission Risks
AI agents frequently operate through service accounts, application identities, API credentials, and delegated user permissions. These identities allow agents to interact with claims management platforms, policy databases, customer records, payment applications, and external services.
However, when an agent inherits broader permissions than its assigned responsibilities require, the potential impact of a compromised identity can extend across multiple enterprise systems.
Real-world scenario:
An insurance company deploys an AI claims-processing agent to retrieve claim documents, update case information, and prepare settlement recommendations.
The agent uses an application identity that has access to several internal systems. While the intended task requires access to a limited set of claims, the same identity can also retrieve unrelated customer records and access additional financial resources. If the agent's credentials are compromised, the attacker may potentially inherit the access available through that identity.
Insurance companies should examine:
- Which user or application identity an AI agent acts as.
- What permissions the identity inherits.
- Which sensitive resources the agent can reach.
- Whether access extends beyond its intended business function.
- How far the potential impact could spread through connected systems.
Grafyn's Agentic Identity Blast Radius Analysis maps the identities agents act as, their permissions, and the resources reachable through those access relationships.
3. Malicious Instructions in Claims Documents and External Content
AI-powered insurance workflows process claim forms, medical reports, repair estimates, customer emails, policy documents, and other external submissions. These materials may contain hidden or malicious instructions that could influence an AI agent's responses or interactions with connected tools.
Real-world scenario:
An insurance company uses an AI assistant to review accident reports and prepare claim summaries. A submitted document contains hidden instructions designed to influence the assistant to retrieve additional policyholder information or interact with a connected claims-management tool beyond its intended purpose.
If the agent cannot adequately distinguish trusted instructions from untrusted document content, the workflow may be exposed to manipulation. The potential impact depends on the agent's permissions, connected tools, safeguards, and available actions.
Insurers should validate external content, restrict tool permissions, maintain clear trust boundaries, and test document-processing workflows against realistic attack scenarios.
Grafyn's Agentic Red Teaming simulates attacks against AI agents and connected workflows to reveal potential attack paths, understand how risks could spread, and validate security controls before deployment.
4. Insecure Third-Party AI Integrations and MCP Servers
Insurance companies depend on external technology providers for document verification, medical data processing, claims assessment, fraud detection, customer support, and analytics. AI agents may connect to these services through APIs, plugins, and Model Context Protocol (MCP) servers.
Every integration introduces additional permissions, credentials, and trust relationships. Poorly secured integrations may expose internal resources through missing authentication, excessive permissions, insecure credentials, or untrusted tool responses.
Real-world scenario:
An insurer connects an AI claims assistant to an external document-processing service through an MCP server. The integration has access to a broad internal claims repository, although the external service only needs specific claim documents.
If the integration is compromised or misused, its permissions could provide a route to information beyond the intended processing requirements.
Security teams should inventory third-party integrations, verify authentication, review permissions, and understand which internal and external systems each connection can reach.
Grafyn's AI Governance helps organizations identify and evaluate MCP security risks through policies that flag issues such as MCP servers without authentication and unapproved or high-risk AI assets. Its Agentic Identity Blast Radius Analysis helps security teams understand the identities, permissions, sensitive data access, and trust boundaries associated with AI agents and their connected integrations.
Read more - MCP Security: Risks, Permissions and Enterprise Best Practices
5. Unauthorized Actions in AI-Powered Claims Automation
AI agents are increasingly capable of performing actions rather than simply generating recommendations. They may retrieve customer information, update claims, route cases, prepare settlement documentation, and interact with connected enterprise applications.
The risk increases when agents can execute consequential actions without appropriate authorization and human review.
Real-world scenario:
An automobile insurance company uses an AI agent to analyze accident photographs, review repair estimates, and prepare settlement recommendations.
The agent is also connected to the claims platform and has write permissions.
If manipulated or provided with misleading information, it may attempt to update claim records or trigger downstream processes outside the expected approval workflow.
The organization must distinguish between read-only access, record modification, and transaction permissions. High-impact actions should require appropriate authorization, validation, human approval, and audit trails.
6. Poisoned Knowledge and Hidden Instructions in Insurance AI Data Sources
Insurance AI systems rely on internal policy documents, coverage guidelines, claims records, document repositories, vector databases, and retrieval-augmented generation (RAG) pipelines to generate responses and support claims decisions. If these knowledge sources contain manipulated documents, inaccurate information, or hidden malicious instructions, AI agents may retrieve and act on misleading content.
Real-world scenario:
Consider an insurance company using an AI-powered claims assistant that retrieves policy documents and coverage guidelines from an internal knowledge repository. An attacker introduces a poisoned document containing inaccurate coverage instructions or hidden malicious content designed to influence the assistant's responses.
When the claims assistant retrieves this document, it may interpret the manipulated information as a legitimate coverage guideline and recommend approving a fraudulent claim that should otherwise be rejected.
The risk can extend further when the claims assistant passes its output to other connected agents responsible for claim verification, eligibility checks, or payment processing. These downstream agents may rely on the incorrect recommendation, allowing the original manipulation to influence multiple stages of the claims pipeline.
Insurance companies should validate knowledge sources, review document integrity, identify hidden instructions, and assess how manipulated content could influence connected AI workflows.
Grafyn's RAG Pipeline Poisoning capability helps organizations assess risks involving manipulated documents, hidden instructions, and compromised content across RAG pipelines and connected knowledge sources. This helps security teams understand how poisoned information could influence AI responses and propagate through connected workflows.
7. Shadow AI Across Insurance Departments
Employees may adopt public AI applications, browser extensions, coding assistants, and third-party agents without organizational approval.
For example:
- Claims employees may upload accident reports or customer documents to public AI chatbots for summarization.
- Underwriting teams may use unapproved AI assistants to analyze applicant information.
- Customer service employees may paste policyholder details into external tools to generate responses.
- Technology teams may install AI coding assistants or browser extensions without security review.
These activities can introduce data exposure risks because security teams may not know which applications process sensitive insurance information.
Real-world scenario:
A claims employee uses an unapproved public AI assistant to summarize a policyholder's medical report. The employee shares sensitive information to complete the task more quickly, but the organization lacks visibility into the application and its associated data exposure risks.
Without AI discovery and usage governance, such applications can remain outside the organization's security oversight.
Grafyn's Shadow AI capability helps organizations discover unsanctioned AI applications, browser extensions, and agents across endpoints, cloud environments, and code. This gives insurance security teams better visibility into unauthorized AI usage and helps identify potential risks involving sensitive policyholder information.
Read more - Shadow AI: What It Is, Risks, Examples, Detection & Prevention
8. Weak AI Governance and Regulatory Accountability
AI systems may support underwriting, pricing, claims handling, fraud detection, and customer interactions. As insurers increasingly deploy AI assistants and agents across business operations, they need visibility into which AI systems are deployed, who owns them, what data they can access, and whether appropriate security policies are enforced.
The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, outlines expectations for insurers' AI governance, risk management, documentation, and oversight.
The NAIC emphasizes that insurers remain responsible for complying with applicable insurance laws and regulations when decisions are made or supported by AI.
Real-world scenario:
Consider an insurance company deploying multiple AI agents to support underwriting and claims operations. An underwriting assistant retrieves applicant information from internal databases, accesses external data sources, and shares relevant information with other connected agents to support application processing.
However, the organization may lack a complete inventory of these AI agents, clear ownership records, or visibility into the identities and permissions they inherit. Some agents may access sensitive applicant information or connect to systems beyond their intended business requirements. Without clearly defined governance policies and enforcement, the organization may struggle to identify unauthorized AI deployments, assess data access risks, or establish accountability for agent activities.
Insurance companies should establish clear AI ownership, maintain inventories of deployed applications and agents, understand their identity relationships and data access, enforce security policies, and implement appropriate human oversight.
Grafyn's AI Governance capabilities help organizations establish visibility into AI assets, define and enforce policies for AI agents and connected infrastructure, and identify policy violations such as unapproved or high-risk agents and MCP servers operating without authentication.
Best Practices for Securing AI in Insurance Companies
Insurance companies need a structured approach to secure AI applications, agents, policyholder data, and connected systems.
Key practices include:
- Discover AI assets: Identify AI applications, agents, copilots, browser extensions, and MCP servers.
- Protect sensitive data: Understand which AI systems can access policyholder, medical, financial, and claims information.
- Enforce least privilege: Restrict agent identities, permissions, and integrations to necessary resources.
- Secure integrations: Review third-party AI services, APIs, and MCP servers for security weaknesses.
- Validate AI outputs: Review important claims, underwriting, and policy recommendations.
- Test AI workflows: Simulate attacks to identify potential exposure and unauthorized access.
- Strengthen governance: Establish AI policies, ownership, accountability, and human oversight.
- Review continuously: Reassess AI assets, permissions, and integrations as systems evolve.
Insurance companies can also use the NIST AI Risk Management Framework (AI RMF) to structure AI risk management.
The Grafyn AI Security Platform helps insurance companies strengthen AI security through AI Discovery & Visibility, Sensitive Data Disclosure, Agentic Identity Blast Radius Analysis, AI Governance, Agentic Red Teaming and Rag Poisioning. Together, these capabilities help insurers improve visibility, understand AI-related exposure, and strengthen security across claims, underwriting, and customer service workflows.
Practical Example: Securing an AI-Powered Insurance Claims Workflow
Consider a health insurance company using an AI claims assistant to review submitted documents, summarize policyholder information, and prepare claim assessments.
The assistant connects to three systems:
- An internal claims management platform.
- A document repository containing medical reports and policyholder information.
- An external document-processing service connected through an MCP server.
The assistant operates through an application identity that allows it to retrieve information and update claim records.
During a security assessment, the company discovers several potential security issues:
- The AI assistant can access medical records beyond those required for its assigned claims.
- Its application identity has permissions to reach unrelated customer information.
- The external document-processing integration has a trust relationship that has not been fully reviewed.
- The agent's connected tools introduce potential pathways for unauthorized actions.
Grafyn can help the security team investigate these relationships by:
- Discovering the AI claims assistant and its connected integrations.
- Identifying sensitive policyholder information accessible through the workflow.
- Mapping the agent's identity, permissions, and reachable resources.
- Identifying trust-boundary crossings involving external services.
- Running attack simulations to assess whether weaknesses could expose additional systems.
- Evaluating governance policies for unapproved agents and insecure MCP servers.
This gives the insurer a clearer understanding of the security relationships surrounding its claims workflow and helps identify potential exposure paths that require attention.
Conclusion:
Artificial intelligence is creating new opportunities for insurance companies to improve claims processing, underwriting, fraud detection, risk assessment, and customer service. However, the growing use of AI agents and connected enterprise systems also expands the potential attack surface.
Policyholder data exposure, excessive agent permissions, prompt injection, insecure third-party integrations, Shadow AI, and weak governance can affect multiple insurance operations.
The challenge for insurers is no longer limited to securing individual AI applications. They must understand the relationships between AI agents, identities, permissions, sensitive information, external integrations, and enterprise systems.
With Grafyn, insurance security teams can discover AI assets, assess sensitive data exposure, analyze agent identity relationships, identify trust-boundary risks, run agentic red teaming simulations, and evaluate AI governance policies.
Secure your insurance AI ecosystem before an exposure becomes a business risk.
Explore the Grafyn AI Security Platform or schedule a demo to understand how Grafyn helps organizations secure enterprise AI environments.



