The prevailing enterprise risk narrative pushed aggressively by some vendors suggests that artificial intelligence has fundamentally altered the cyber-threat landscape, drastically accelerating breach velocity, driving up incident costs, and exposing organizations to existential threat vectors. Senior leadership teams are now being routinely urged to deploy specialized AI-focused defensive technology to counter these emerging risks.

I was reading a report published by a vendor recently which spoke about how:

·      AI is driving up the cost of cyber breaches in 2026 to unprecedented levels

·      Year-on-year, the average breach cost rose by an “alarming factor” in 2025  

·      Over a quarter of the attacks studied by them involved AI

·      AI-driven attacks are now getting faster and more expensive to fix

While there were references to phishing and other modes of attacks, the role of AI as an enabler is not at all highlighted. Further, rigorous evaluation of the available empirical data reveals a critical disconnect between public risk perception and underlying operational reality. Artificial intelligence is undeniably being utilized by threat actors to optimize conventional execution and in localized social engineering and automated reconnaissance, the immediate operational crisis facing most organizations stems not from novel, ultra-sophisticated AI attacks, but from systemic governance failures, unmanaged identities, unencrypted sensitive repositories, and shadow AI deployments. Lack of AI guardrails and standards don’t help either.

Lastly, opaque and vague cost of breach calculations dilute the perceived integrity of such reports.

Headline figures asserting massive cost premiums for "AI-driven breaches" relies heavily on survey-based methodologies, activity-based cost modeling, and correlation-heavy data sets that often fail to establish direct causation or control for organizational size, security maturity, and incident severity. In today’s blog post, we conduct an assessment that interrogates the empirical evidence base surrounding AI cyber risk, disentangles correlation from causation in breach economics, highlights acute regional data gaps, and outlines an evidence-led control hierarchy for CISOs, boards, and investment committees.

The AI cyber-risk narrative: What we actually know

To evaluate enterprise exposure accurately, executive leadership must separate proven operational facts from analytical inferences and technology-market framing.

Empirical foundations vs. market assumptions

• Assisted execution vs. autonomous offense: Current evidence demonstrates that generative AI acts primarily as an efficiency multiplier for conventional attack vectors across the cyber kill chain. It enables threat actors to generate highly believable, localized phishing, voice deepfakes, and automated scanning scripts at scale. Available evidence does not prove that autonomous, self-propagating AI agents are routinely conducting end-to-end cyber intrusions without human intervention.

• The "AI Breach" definition ambiguity: Industry studies frequently conflate four distinct risk surfaces under a single "AI-related breach" label:

  • Attackers using AI as an auxiliary efficiency tool (such as writing a phishing template).

  • Attacks specifically targeting AI models or infrastructure (such as prompt injection, model inversion).

  • Traditional breaches occurring in environments where unapproved AI tools ("shadow AI") were present.

  • Standard IT failures or cloud misconfigurations affecting workloads that happen to host AI applications.

• The governance vulnerability: The vast majority of reported security incidents involving enterprise AI workloads trace back to fundamental architectural and operational hygiene failures. This includes exposed APIs, misconfigured cloud storage, excessive privileges, and missing access controls. 

Methodological and evidentiary limitations of industry studies

Much of the executive consensus regarding cyber-breach economics relies on annual benchmark studies commissioned (and/or influenced) by technology vendors and conducted by private research entities. While these publications offer valuable directional insights, a more realistic advisory perspective requires pointing out their severe methodological constraints before using them to justify capital allocation.

• Non-statistical sampling: Leading breach studies frequently draw from non-statistical, judgmental samples. Consequently, margins of error, standard deviations, and confidence intervals cannot be calculated. Applying strict statistical significance tests to these datasets is virtually impossible.

• Unverified self-reporting and recall bias: Data is largely gathered through post-incident qualitative interviews with executive and IT personnel. Findings reflect the subjective recollections, memory limitations, and internal framing of respondents rather than independently verified forensic telemetry or audited financial general ledgers.

• Lack of actual incident or event information: Such reports while talking about huge rise in AI-driven cyberattacks fail to cite a single case of an actual event reported by a victim they surveyed

• Modeled vs. accounting costs: Financial figures presented in industry benchmark reports do not represent direct general ledger accounting losses. Instead, they are constructed using activity-based cost (ABC) models that extrapolate indirect expenses, operational friction, estimated downtime impact, and projected customer churn. The numbers are not reported by the entity concerned but calculated by the survey vendor who may or may not be able to perceive the actual cost of a dataset lost by a victim.  

• Omitted variables and maturity bias: Research sample frames tend to be biased toward medium-to-large enterprises with formal security programs. Crucially, these studies rarely employ multivariate regressions to isolate AI usage from confounding variables such as organizational complexity, total IT budget, regulatory footprint, or baseline security maturity.

The inherent problem with AI-driven breach economics

Headline figures often assert that AI-driven breaches impose massive direct financial penalties, often citing an additional cost of ~$1 million per incident, or that defensive Security AI saves millions. These economic claims require rigorous critical evaluation. They should be grounded in a clear articulation of the variables considered with long term and short term implications called out with assumptions being cited as well.  

Deconstructing the AI Cost Penalty

When research indicates that attacks utilizing AI cost substantially more than conventional malicious breaches, risk committees must ask whether AI is the direct cause of the cost inflation or merely a correlated variable.

• Targeting high-value entities (HVEs): Advanced threat actors employing sophisticated AI-assisted tooling overwhelmingly target large, highly complex, and lucrative enterprises (such as global financial institutions, energy grids). These target organizations naturally incur higher breach recovery, legal, regulatory, and downtime costs regardless of the threat actor's specific tooling.

• Incident severity and scope: An attack that leverages automation to move faster may compromise a larger footprint of records or operational technology before containment. The elevated cost is a function of the scope of destruction and recovery downtime, not the mathematical presence of an AI script.

• Absence of variance and sample sizes: Industry datasets generally do not disclose the exact sample size () of incidents categorized purely as "AI-driven," nor do they report the variance or standard deviation across those incidents. Without these metrics, asserting a definitive, causal economic baseline is methodologically unsound.

The defensive AI cost-reduction paradox

Conversely, claims that extensive deployment of security AI and automation yields drastic cost reductions (such as savings approaching $2 million per breach) must be viewed through the lens of overall security maturity:

• Adoption as a proxy for maturity: Organizations that are capable of extensively deploying security AI across prevention, detection, and response workflows almost universally possess higher cybersecurity budgets, mature Security Operations Centers (SOCs), robust Identity and Access Management (IAM), well-rehearsed Incident Response (IR) plans, and formal DevSecOps practices.

• Causation fallacy: The lower breach cost observed in high-AI-adoption organizations is driven by their comprehensive, baseline security maturity. Attributing these financial savings solely to the purchase of AI security tooling represents a fundamental failure to control for confounding operational variables.

The India evidence gap

When assessing enterprise cyber risk in India, boards and CISOs must establish a strict line between global statistical generalizations and verified, localized operational evidence.

Critical assessment of local data

• Sample Size Limitations: Global data breach studies typically include a limited cohort of Indian enterprises (~50–60 organizations, representing under 10% of global samples).

• Unsubstantiated specifics: Available global research datasets do not establish or disclose:

  • The precise number of Indian breaches classified strictly as "AI-driven."

  • The prevalence of specific AI attack vectors in India (e.g., model inversion, prompt injection, data poisoning).

  • Independently verified attack chains, malware signatures, or forensic evidence proving AI was essential to localized intrusions.

  • Verified accounting breakdowns isolating AI-attributable losses within Indian entities.

Local operational context vs. Global framing

In India's financial and critical infrastructure sectors, the practical threat landscape remains dominated by high-volume digital financial fraud, voice/SMS phishing (vishing/smishing), credential harvesting, and social engineering aimed at remote account takeover. While threat actors utilize basic generative tools to improve the linguistic quality and volume of localized phishing, treating this as a fundamentally new category of "AI-driven cyber breach" is a motivated misrepresentation of operational reality.

Taxonomy of AI cyber risk: Disentangling threat vectors

To prevent generic framing and improper resource allocation, organizations must separate AI cyber risk into four distinct categories:

A. AI-Assisted conventional attacks

• Definition: Traditional attack methods where attackers use standard commercial AI tools as auxiliary productivity aids to save time.

• Examples: Writing convincing phishing emails, translating lure messages, refining reconnaissance scripts, or formatting social engineering personas.

• Impact: Lowers the barrier to entry and increases attack volume, but does not alter the underlying vulnerability or intrusion vector.

B. AI-enabled high-velocity attacks

• Definition: Attacks where specialized AI or automation materially alters the speed, scale, targeting precision, or capability of the exploit payload.

• Examples: Automated vulnerability discovery and rapid exploit synthesis, polymorphic AI-generated malware payload obfuscation.

• Impact: Compresses defender reaction windows from days to minutes.

C. Attacks targeting AI systems directly

• Definition: Exploits directed at the internal mechanics, architecture, or training pipelines of AI models and applications.

• Examples: Indirect prompt injection, model inversion/extraction, training data poisoning, adversarial evasion, and compromise of connected model APIs or orchestration plugins.

• Impact: Compromises data confidentiality, model integrity, and system decision-making logic.

D. AI-related security and governance failures

• Definition: Incident execution resulting from poor management, weak identity architecture, or missing oversight surrounding AI deployments.

• Examples: Deployment of unapproved third-party AI software ("Shadow AI"), excessive API service account permissions, lack of data classification, and storing unencrypted sensitive data in public cloud AI repositories.

• Impact: Represents basic operational hygiene failure, exposing sensitive assets without requiring sophisticated attacker techniques.

Real-world attack scenarios vs. Evidence gaps

A foundational risk assessment must evaluate whether published industry claims are backed by complete, forensically verified end-to-end attack chains.

Across broader industry literature, there is a clear scarcity of documented, real-world case studies detailing end-to-end forensic attack chains where an AI model's mathematical failure was the primary breach vector.

Instead, the overwhelming majority of verified incidents trace back to standard initial access vectors: voice/SMS phishing, abused valid credentials, unpatched public-facing software vulnerabilities, and third-party software supply chain compromises.

Re-examining financial-services cyber risk

Financial institutions are routinely cited as the primary target for advanced AI cyber threats. A nuanced examination reveals important distinctions:

• Broad industry aggregation: Benchmark studies routinely aggregate diverse sub-sectors such as retail banking, investment banking, non-banking financial companies (NBFCs), payment processors, capital markets, digital lending platforms, and cryptocurrency exchanges into a single "Financial Services" bucket.

• Misinterpreting fraud as cyber intrusions: A significant proportion of reported financial sector incidents involve consumer-facing payment fraud, social engineering, credential harvesting, and authorized push payment scams.

• Regulatory cost amplification: Financial services face stringent global data protection regulations, mandatory breach notification rules, and high compliance penalties. Consequently, when a financial institution experiences a breach, statutory reporting costs and customer compensation mechanisms skew total breach cost figures upward—regardless of whether AI was involved in the intrusion.

Governance vs. technology: The true operational bottleneck

The primary vulnerability facing modern enterprises is not a lack of sophisticated AI security defense technology; it is the rapid adoption of enterprise AI systems in the absence of basic governance, identity management, and operational hygiene.

The governance breakdown

• Industry benchmark data demonstrates that over two-thirds of breached organizations lacked functional AI governance policies or shadow-AI detection capabilities.

• Furthermore, over 90% of organizations suffering security incidents involving AI models failed to enforce fundamental access controls (such as role-based access control and multi-factor authentication) over those AI workloads.

• Fewer than half of enterprise environments actively manage and secure Non-Human Identities (NHIs)—such as API keys, service accounts, and automated tokens—operating within automated AI workflows.

Enterprise control hierarchy

Organizations attempting to mitigate AI risk by immediately purchasing specialized AI security platforms are jumping to Step 8 of the defense hierarchy while ignoring Steps 1 through 7.

Executive leadership must enforce a strict sequence of controls:

• Governance and policy frameworks: Define clear approval workflows, acceptable use policies, and compliance oversight for AI tools.

• Data classification and encryption: Enforce robust encryption at rest and in motion across all enterprise repositories prior to exposing data to AI pipelines.

• Identity & Access Management (IAM): Enforce strict least-privilege policies, Zero Trust access, and dedicated lifecycle management for both human users and Non-Human Identities (NHIs).

• Visibility and shadow-AI discovery: Implement active discovery tools to detect unauthorized cloud workloads, unapproved SaaS AI usage, and rogue API integrations.

• Basic hygiene and DevSecOps: Maintain disciplined vulnerability management, continuous patching, and automated configuration auditing.

• Advanced AI tooling: Deploy specialized AI security monitoring only after basic access, identity, and data controls are fully operational.

The technology-commercial narrative

When evaluating the growing pressure to procure dedicated AI cybersecurity solutions, investment committees must consider the broader commercial dynamics driving the vendor ecosystem:

• Commercial Incentives: Cybersecurity product vendors have strong commercial incentives to frame emerging risks around novel, complex threats that necessitate new, specialized software suites.

• Defensive Tool Proliferation vs. Fundamental Controls: Marketing narratives often overemphasize frontier AI threats while downplaying the reality that over half of enterprise data breaches result from basic, unencrypted sensitive repositories, credential abuse, or standard phishing.

• Maturity Proxy Fallacy: Commercial messaging frequently implies that purchasing AI defense systems directly drives lower breach costs. In reality, lower breach costs are achieved through holistic cybersecurity program maturity, strong operational discipline, well-funded SOCs, and effective governance.

Core lessons: Cybersecurity fundamentals first

A critical evaluation of the evidence leads to an undeniable conclusion: The effective mitigation of AI-related cyber risk relies primarily on mastered cybersecurity fundamentals, not specialized vendor solutions.

The primary factors proven to consistently mitigate breach impact center on fundamental disciplines:

• DevSecOps Approaches: Embedding security testing directly into development lifecycles remains one of the top factors reducing incident costs.

• Comprehensive Identity Security: Rigorous IAM implementation across human and machine identities prevents unauthorized lateral movement.

• Universal Data Encryption: Protecting data at rest and in motion minimizes financial exposure, even when network perimeters are breached.

• Incident Response Readiness and Testing: Regular exercise of incident response plans, red teaming, and threat hunting drastically compresses breach containment timelines.

Analytical evidence matrices

Global Matrix: What the Evidence Proves vs. What It Doesn't

India-specific evidence matrix

Strategic recommendations for CISOs, boards, and investment committees

• Audit baseline governance before software procurement:

Mandate clear governance policies defining approved AI deployment frameworks, data handling rules, and mandatory approval gates prior to authorizing capital expenditure for AI security products

• Extend identity architecture to Non-Human Identities (NHIs):

Treat API keys, service accounts, automated tokens, and AI agents as high-risk identity vectors. Implement continuous runtime authorization, automated key rotations, and strict role-based access controls

• Enforce universal encryption standards:

Ensure that sensitive customer PII, employee records, and corporate intellectual property are encrypted both at rest and in transit  Encryption remains the single most dependable control to neutralize data exposure during an intrusion

• Maintain continuous visibility over Shadow AI and APIs:

Deploy network and endpoint discovery mechanisms to monitor unapproved SaaS AI tools, unauthorized cloud storage instances, and shadow API integrations

• Prioritize DevSecOps and operational readiness:

Allocate capital toward baseline security practices—such as integrating security into development lifecycles, rigorous vulnerability management, regular red-teaming exercises, and rehearsing incident response plans

Conclusion

The central challenge facing modern enterprises is not that cyber adversaries have unlocked unstoppable, autonomous AI weapons. Rather, it is that enterprises are adopting and scaling artificial intelligence far faster than they can govern, secure, and monitor it. Some are also not factoring AI-linked risks in their risk calculations.  

To mitigate cyber risk effectively, CISOs, board members, and investment committees must resist vendor-driven hype and retain analytical rigor. The most impactful response to AI-related cyber risk does not begin with acquiring expensive, unproven AI defense software. It begins by mastering foundational security disciplines: enforcing strict governance, securing human and machine identities, encrypting sensitive repositories, and maintaining relentless operational hygiene across the entire enterprise estate.

Download this white paper on AI in OT governance to learn more.