SecurityBrief Asia - Technology news for CISOs & cybersecurity decision-makers
Asia
Mainframe security after the perimeter: moving from inherited trust to continuous assurance

Mainframe security after the perimeter: moving from inherited trust to continuous assurance

Thu, 19th Nov 2026 (Today)
Kaushik Bagchi
KAUSHIK BAGCHI General Manager - Asia Pacific Rocket Software

The mainframe is rarely the weakest component in the enterprise technology estate. The greater risk increasingly sits in the connections around it: privileged identities, host-access channels, APIs, replicated data, software supply chains and AI agents. Security leaders should therefore treat the mainframe as part of a hybrid control system and demand continuous evidence that controls remain effective as the environment changes.

A strong platform can still sit inside a weak system

For decades, mainframe security has benefited from disciplined access controls, tightly managed change processes and a culture of operational resilience. Those strengths remain important. They are not, however, a substitute for examining the wider system through which people, applications and machines reach mainframe applications and data.

The attack surface is expanding in four directions. First, modernisation exposes transactions through APIs and integration platforms. Second, data is copied or streamed into cloud analytics, customer-service and AI environments. Third, remote administration and terminal access create identity-dependent pathways into sensitive workloads. Fourth, AI agents are beginning to investigate, recommend and eventually initiate operational actions across multiple tools.

Industry breach data reinforces the need to focus on these pathways and the environment in which mainframes now operate. Verizon's 2025 Data Breach Investigations Report found that exploitation of vulnerabilities accounted for 20% of breaches, close to credential abuse at 22%. In its Cost of a Data Breach Report 2025, IBM warned that organisations are adopting AI faster than they are implementing governance, and reported average ransomware-related breach costs of USD 5.08 million.

The control objective is assurance, not simply compliance

Traditional mainframe reviews often concentrate on configuration, entitlements, segregation of duties and evidence collected at a point in time. These remain necessary, but they can create false confidence if they do not test executable code, control drift, patch exposure, access pathways and the behaviour of connected applications.

A continuous-assurance model asks a different set of questions:

  • What has changed in the operating environment since the last assessment?
  • Which vulnerabilities are realistically exploitable in the context of existing privileges and workloads?
  • Are compensating controls operating as designed, and can that be demonstrated?
  • Can critical data be located across copies, extracts, test environments and analytics platforms?
  • Can every human or machine action be tied to an approved identity, policy decision and auditable outcome?

This approach helps organisations find and fix security risks faster, verify that issues are resolved, and clearly understand what risk remains.

Four risk scenarios that keep IT leaders awake at night 

1. A compliant configuration conceals vulnerable authorised code

Illustrative risk: A financial institution completes its annual IBM z/OS configuration review with no material exceptions. A third-party module nevertheless contains an integrity exposure that executes with elevated authority. The issue is not visible to a configuration-only assessment.

Control response: Add binary and runtime-aware vulnerability analysis to configuration review. Establish a baseline, validate findings with the software vendor, patch or mitigate, and rescan. Link technical severity to the affected business service, privilege level and data classification.

2. Identity compromise enters through the host-access layer

Illustrative risk: A contractor's credentials are captured through phishing and reused through a legacy terminal-access path. The mainframe has not been directly "hacked"; an authorised route has been used by an unauthorised actor.

Control response: Integrate host access with enterprise identity providers, phishing-resistant MFA or passwordless authentication, device and session context, least privilege and session audit. Treat the terminal path as part of the zero-trust architecture, not as a trusted internal channel.

3. Modernisation replicates sensitive data faster than governance can follow

Illustrative risk: A customer-service programme streams Db2, IMS, VSAM or ADABAS data into cloud analytics and an AI assistant. Copies proliferate across development, testing and vector-search environments. Privacy, retention and deletion obligations become difficult to execute consistently.

Control response: Map lineage and ownership before movement. Minimise copies, mask or tokenise non-production data, enforce purpose-based access and retain data at source where practical. Test whether privacy requests and retention policies can be executed across both systems of record and downstream derivatives.

4. An AI agent receives more authority than the task requires

Illustrative risk: An operations agent investigates a CICS performance issue across logs, dashboards and automation tools. A long-lived service credential allows broader access than the investigation needs, and the resulting action is difficult to reconstruct.

Control response: Evaluate policy at invocation time, not only at account provisioning. Issue short-lived, task-scoped credentials; require human approval for material actions; record the request, evidence, policy decision, identity and outcome; and prevent unsupervised production changes until control effectiveness is proven.

Agentic AI changes the identity and audit model

AI used for mainframe operations should be treated as a new class of privileged actor. Unlike a conventional service account, an agent may select tools, assemble evidence and adapt a sequence of steps dynamically. That makes static role assignment insufficient on its own.

A sound control architecture should separate reasoning from authority. The agent may investigate and recommend, while a policy enforcement layer decides whether a specific tool invocation is permitted. Where execution is approved, the identity should be temporary, narrowly scoped and revoked when the task ends. This direction is consistent with NIST's zero-trust principle that trust should not be implied by network location or asset ownership.

Emerging products illustrate this pattern. Rocket EVA 2.0, for example, is positioned as an investigation platform that correlates evidence across mainframe and hybrid operations. Its PlanGuard capability evaluates the caller, request, tool and context before execution, then creates a task-specific credential and records the decision trail. The design is notable less as a product feature than as an example of the controls organisations will increasingly require for any operational AI agent.

Technology should support the control model, not define it

The market is beginning to converge around several capability categories. Organisations should evaluate them against their target control architecture and existing investments rather than pursue a single-vendor security stack.

These solutions are focusing on more dynamic governance and control covering areas including mainframe vulnerability assurance; host and privileged access management; application security analysis and controls during modernisation; data security and privacy controls; and AI-agent governance. 

A pragmatic 12-month assurance agenda

0–90 days | Establish exposure and ownership

Define the critical IBM Z services, map privileged identities and access paths, inventory cryptographic assets and connected data flows, and agree on the risk taxonomy. Run targeted vulnerability and access-path assessments on the highest-impact services.

3–6 months | Integrate security into modernisation

Add application dependency analysis, security tests and data controls to API, cloud and DevOps initiatives. Remediate high-risk access paths, implement stronger authentication and establish recurring vulnerability and patch validation.

6–12 months | Operationalise continuous evidence

Automate control monitoring, connect findings to business services, track closure through rescanning, and produce risk dashboards for technology and risk committees. Pilot AI-assisted investigation only with invocation-time policy, scoped credentials and human approval.

What good looks like

Boards and executives do not need another dashboard of raw vulnerabilities. They need evidence that material exposures are identified quickly, prioritised consistently and closed within an agreed risk appetite. Useful measures include median time from exposure to validation, percentage of critical findings with an identified owner, patch and mitigation ageing, rescan success, privileged-access anomalies, control drift, and the proportion of critical services with current assurance evidence.

The strategic implication is straightforward: mainframe security should no longer be considered a specialist exception operating behind a trusted perimeter. It is a core component of enterprise cyber resilience. The organisations best positioned to modernise will be those that preserve the mainframe's disciplined control culture while extending it across APIs, data platforms, access channels and AI-driven operations.