The Double-Edged Sword of AI & Cyber Risk

Siddhartha Dalal & Farhan Shah

Professor, Columbia University · Advisor, Trivora — Founder & CEO, Trivora

A practical underwriting framework for MGAs

AI, software vulnerability discovery, and the coming tidal wave of SMB cyber claims.

Abstract

Artificial intelligence is changing cyber risk in two related ways. First, AI-assisted cyber reasoning systems can help defenders find and patch software vulnerabilities faster than traditional manual review. The same capability, if misused or leaked, can also compress the time between vulnerability discovery and exploitation. Second, AI coding assistants are increasing the volume of business code, integration scripts, dashboards, and automation written by small and mid-sized businesses (SMBs). This raises a practical underwriting concern: SMBs may deploy more code, faster, with weaker review, weaker identity controls, and incomplete evidence that the code is secure. It may result in a tidal wave of claims. This paper explains the risk for an SMB insurance audience, separates documented developments from forward-looking scenarios, and proposes underwriting questions, policy conditions, security controls, and a simple premium adjustment framework.

01Executive summary for MGA leaders

The insurance issue is not simply that AI makes attackers more dangerous. The larger issue is that AI changes the speed, scale, and distribution of software risk.

AI-assisted vulnerability discovery can shorten the time required to find serious software defects. This may increase claim frequency if remediation lags behind discovery.
SMBs are especially exposed because they often rely on vendors, managed service providers, SaaS tools, and informal “glue code” connecting business systems.
AI-generated code can expand the attack surface faster than traditional security review processes can keep up.
Underwriters should ask not only “Do you use AI?” but also “Where does AI touch production systems, sensitive data, credentials, code, and vendor integrations?”
The most practical MGA response is a combination of underwriting questions, minimum controls, targeted exclusions or sublimits, security credits, and incident-response triggers.

02Scope and important caveat

This paper addresses two categories of AI-related cyber insurance risk:

Cyber reasoning and vulnerability discovery models: AI systems that assist in finding, validating, prioritizing, or patching software vulnerabilities.
AI-generated code: code, scripts, workflows, dashboards, API connectors, and automation created or substantially modified using AI coding tools.

Some examples in this document are forward-looking scenarios. They should be treated as underwriting stress cases.

03Risks related to cyber reasoning models

What cyber reasoning systems do

A cyber reasoning system like Mythos and GPT-5.3-cyber are AI-assisted tool that can help identify, triage, exploit-test, or patch software vulnerabilities. These systems are already being explored in public and government-backed settings. For example, DARPA’s AI Cyber Challenge focused on autonomous systems that can find and patch vulnerabilities in open-source software used in critical infrastructure.

For defenders, this is positive: serious vulnerabilities can be found earlier and patched faster. For insurers, the concern is timing. If many vulnerabilities are discovered at once, policyholders may face a surge of patching work, temporary operational disruption, and heightened exposure while fixes are being tested and deployed.

Why this matters more for SMBs

Large enterprises usually have security engineering teams, vulnerability management programs, test environments, and formal change control. Many SMBs do not. They depend on:

software vendors to issue patches;
managed service providers to apply them;
SaaS platforms and plug-ins;
informal scripts and integrations built by employees or consultants;
backups and identity controls that may not be regularly tested.

This means an SMB can be exposed in three ways at the same time: vendor software may require urgent patching, internal integrations may break during remediation, and attackers may move faster than the business can respond.

Main insurance implications

For an MGA, the underwriting question is not whether AI exists in the market. It is whether the insured can absorb a faster vulnerability cycle. Relevant claim pathways include:

ransomware following delayed patching;
business interruption during emergency remediation;
data breach caused by exposed integrations or credentials;
dependent business interruption if a core SaaS or cloud vendor is affected;
regulatory or contractual claims if sensitive data is mishandled.

Regulatory risk may also increase where AI systems or sensitive data are involved. Under the EU AI Act, certain violations can carry penalties up to EUR 35 million or 7 percent of worldwide annual turnover, depending on the obligation breached.

04Risks related to AI-generated code

Why AI-generated code is a different risk

AI coding assistants can increase software output dramatically. A small engineering team can now create more scripts, dashboards, API connectors, and internal applications than it could before. This productivity gain also creates a risk: more code may enter production with less design review, less threat modeling, and less testing.

The most important SMB exposure may not be polished commercial software. It may be the “glue code” that connects accounting systems, CRM platforms, customer portals, payment tools, data warehouses, marketing tools, and AI services.

Evidence of security concern

Recent industry research has found that AI-generated code can introduce security flaws in a significant share of coding tasks. Veracode’s 2025 GenAI Code Security Report, for example, reported security vulnerabilities in 45 percent of AI-generated coding tasks it evaluated across more than 100 models. Academic work also indicates that vulnerability rates vary by language, repository context, and tool, so a single percentage should not be treated as universal.

The underwriting conclusion is not that all AI-generated code is unsafe. The conclusion is that AI-generated code requires evidence of review, testing, access control, and ownership.

Typical SMB failure modes

Over-permissioned AI agents: an AI tool or automation account receives administrator privileges because it is easier than configuring least privilege.
Shadow AI: employees paste customer data, contracts, financial records, or health information into unapproved AI tools.
Unreviewed code: AI-generated scripts are copied into production without secure code review.
Credential leakage: API keys, secrets, or tokens are embedded in generated code or shared with third-party tools.
Broken business logic: generated code functions correctly in simple tests but fails under edge cases, leading to data exposure, payment errors, or operational disruption.

05Recommended SMB AI security rider

The following clauses are drafted as underwriting concepts, not final legal wording. They should be reviewed by coverage counsel.

Clause A · Critical vendor security notice

Requirement. The insured must notify the insurer or designated incident-response contact within a defined period, such as 24 hours, after receiving a critical security notice from a core technology vendor or managed service provider.

Purpose. This creates an early warning mechanism when a vendor vulnerability may affect many insureds at once.

Clause B · Immutable recovery capability

Requirement. The insured must maintain immutable or logically air-gapped backups for critical systems. Restoration must be tested at least every 90 days, with evidence retained.

Purpose. Ransomware and destructive attacks are more survivable when restoration is tested before an incident.

Clause C · AI agent identity and least privilege

Requirement. AI agents, automation tools, and integration scripts must run under non-human service accounts with least-privilege permissions. Administrative actions, destructive actions, and bulk data exports should require human approval or compensating controls.

Purpose. This limits the blast radius if an AI agent, script, or integration is misused or compromised.

Clause D · Shadow AI discovery and data handling

Requirement. The insured must maintain an inventory of approved AI tools and third-party AI integrations. Unapproved tools handling sensitive data must be removed, blocked, or formally risk-accepted within a defined period, such as 72 hours after discovery.

Purpose. This addresses the common SMB problem of employees using free or unmanaged AI tools with customer or business data.

Clause E · AI-generated code review

Requirement. Code created or substantially modified by AI must be reviewed before production use. Review should include at least one of the following: secure code review, static application security testing, dependency scanning, secrets scanning, or controlled testing before deployment.

Purpose. This does not prohibit AI-generated code. It requires evidence that the insured has a process to manage it.

06Underwriting questions

MGAs can add a short AI-specific supplement to cyber applications. Recommended questions include:

Does the applicant use AI coding assistants to create production code, scripts, dashboards, or integrations?
What percentage of production code or automation was created or substantially modified using AI in the last 12 months?
Are AI-generated code changes subject to human review before production?
Are secrets, API keys, and credentials scanned before code deployment?
Are AI agents or automation tools restricted to least-privilege service accounts?
Does the applicant maintain an inventory of approved AI tools?
Are employees prohibited from entering sensitive customer data into unapproved AI systems?
Are backups immutable or logically air-gapped, and when was the last successful restore test?
Is there a written procedure for responding to critical vendor security notices?
Does the applicant use a managed detection and response provider, vulnerability scanner, or SaaS security posture management tool?

07Premium adjustment framework

A simple conceptual model is:

Premium = B × (1 + α_AI D_AI) × (1 + β_AI V_AI) − C

B is the base cyber or technology E&O premium.
D_AI is AI code density, meaning the estimated share of code, scripts, or integrations created or substantially modified using AI.
V_AI is observed vulnerability density in AI-generated or AI-assisted code, based on testing history, scan results, or incident experience.
α and β are underwriting sensitivity factors that vary by industry and data sensitivity.
C is the value of security credits for controls such as human review, least privilege, immutable backups, secrets scanning, and continuous monitoring.

These values should be calibrated with claims experience. At the early stage, the model should be treated as a structured underwriting aid, not as a fully credible actuarial model.

08Claims scenarios for stress testing

Scenario 1 · Vendor vulnerability and delayed patch

A core SaaS vendor issues a critical security notice. The SMB does not notify its managed service provider promptly, and patching is delayed. Attackers exploit the vulnerability and deploy ransomware. Coverage questions may involve business interruption, incident-response timing, and whether the insured followed notification and patch-management conditions.

Scenario 2 · AI-generated integration exposes customer data

An employee uses an AI tool to create a script connecting a CRM to a reporting dashboard. The script stores an API key in plain text and permits excessive data export. A threat actor obtains the key and downloads customer records. Coverage questions may involve privacy liability, security failure, professional services, and whether AI-generated code was reviewed.

Scenario 3 · Shadow AI and sensitive data leakage

An employee pastes customer records into an unapproved AI summarization tool. The data includes personally identifiable information. Coverage questions may involve data handling exclusions, regulatory defense costs, notification expenses, and whether the insured maintained an approved AI tools policy.

09Implementation roadmap for an SMB MGA

Add a short AI usage supplement to cyber applications.
Define minimum controls for higher-risk accounts: immutable backups, MFA, least privilege, AI tool inventory, secrets scanning, and code review.
Create security credits for insureds that can provide evidence of controls.
Create referral triggers for high AI code density, AI agents with administrative permissions, or lack of backup testing.
Build an accumulation view for common vendors, AI platforms, MSPs, and SaaS dependencies.
Track claims involving AI-generated code, shadow AI, AI-assisted fraud, and vendor vulnerability events.

10Conclusion

AI will not replace the core questions of cyber underwriting. It will make them more urgent. For SMBs, the greatest risk is not a science-fiction attack. It is ordinary software, ordinary credentials, ordinary SaaS tools, and ordinary employee behavior accelerated by AI. MGAs that can ask better questions, require practical controls, and price based on evidence will be better positioned than those that treat AI as a generic exclusion or a marketing slogan.

References

1. DARPA, “AI Cyber Challenge marks pivotal inflection point for cyber defense,” 2025.

2. EU AI Act Service Desk, Article 99 penalties.

3. Veracode, 2025 GenAI Code Security Report, reported via Business Wire, 2025.

4. Schreiber and Tippe, “Security Vulnerabilities in AI-Generated Code: A Large-Scale Analysis of Public GitHub Repositories,” arXiv, 2025.

Confidential draft for SMB-focused Managing General Agent, Trivora, Inc. Siddhartha Dalal, Professor, Columbia University, Advisor, Trivora, Inc. May 13, 2026.