ARTICLE / 02
AI in Auditing

AI in IT Auditing: Reviewing Vulnerability Tickets Marked as Resolved

How auditors can use AI to identify inconsistencies in vulnerability evidence while preserving source verification, governance, and professional judgment.

Published
Reading time
6 min read
Author
Angélica Rosario Maisonet
↓

Reading vulnerability tickets can take time because a status tells only part of what happened. A ticket may say “resolved” while a comment says the fix has been merged and is waiting for deployment. The auditor then has to check the affected application, the release records, and the organization’s definition of “resolved” to understand whether the issue was handled correctly.

This is a practical use for an AI assistant approved for audit work. It can help review closure notes and related records, summarize explanations, and identify cases that need a closer look. The IIA describes uses of generative AI in audit planning, evidence gathering, reporting, and follow-up. Here, I would use it for a focused task: finding possible differences between what a vulnerability ticket says and what the release records show.

A fix can be at different stages

Suppose a scanner detects a vulnerable third-party dependency. Renovate or Dependabot prepares an update, but the tests fail. A developer adjusts the code, the team reviews it, and the change is approved for a later release. Meanwhile, someone closes the vulnerability ticket with the note “fix complete.”

In this example, Renovate and Dependabot handle dependency updates through rules and package information. That automation alone is not evidence of AI use. The AI use being discussed is the auditor asking an assistant to review the ticket’s explanation alongside scan, approval, and deployment records.

A SQL query can compare the closure date with the deployment date. The notes require more interpretation. “Fixed,” “merged,” “mitigated,” and “risk accepted” describe different situations, and the reason for closing a ticket may appear several comments earlier. An assistant can help locate those explanations and identify wording that seems inconsistent with the timeline. The auditor can then check the exact identifiers and dates in the source systems.

If policy requires deployment and a confirming scan before closure, a ticket closed while the fix was still waiting for release may need further investigation. If policy allows closure after the fix is merged, with the deployment tracked separately, the same sequence may be acceptable. The auditor needs the applicable policy and any approved exception before deciding whether the case supports a finding.

There is also a reporting question. Would someone reading “resolved” understand that the production fix was still pending? Even if the closure followed policy, the status may give management an unclear picture of the remaining risk.

The environment and version matter throughout this review. A fix in a test environment does not establish that production received it. A clean scan of another branch cannot confirm that the affected release was corrected. One ticket may also cover several applications with different deployment dates. The auditor needs to connect each alert to the application, version, and environment being assessed.

Giving the assistant a useful task

Before reviewing tickets with AI, the audit team needs to establish the criteria. The organization’s policy should explain how it handles remediation, accepted risk, compensating controls, and emergency changes. Those criteria should be included in the review so the assistant does not assume that every early closure is a violation.

The team also needs a complete set of cases within the audit’s scope. Vulnerability alerts should be related to affected applications, tickets, and deployment records. Starting with closed tickets alone would leave out alerts that never had a ticket. AI cannot identify those gaps if the relevant records are missing from the data it receives.

I would give the assistant a specific request, such as: “Review the closure notes for these tickets. Identify cases where the note says the fix is complete but the related records show a pending production deployment. Include the ticket ID and the records that support each case. If information is missing, say what is needed.” This gives the auditor something concrete to check.

The work should take place in an approved environment using only the information needed for the task. Code, customer information, credentials, and sensitive vulnerability details need to be handled according to the organization’s rules. The team should also know which records the assistant can access and whether it has permission to perform actions.

Once the assistant returns results, the auditor checks the ticket and release identifiers, compares dates with a repeatable query, and reads the original comments. Questions about an unclear status or exception should go to the control owner. The working papers should retain the criteria, the full population being reviewed, the method, verified evidence, limitations, and the reasoning behind the conclusion.

Cases the assistant did not flag need attention as well. A sample of these records, including known complex or emergency changes, can help identify missed problems. Reviewing only the suggested exceptions would provide little information about what the assistant overlooked.

The IIA’s requirement for relevant, reliable, and sufficient information applies to this work too. The assistant’s explanation needs to lead back to records the auditor can examine. A convincing summary without that support gives the auditor more work to verify before reaching a conclusion.

Checking the assistant’s work

An assistant may connect a ticket to the wrong release, miss a relevant comment, or refer to a scan that never happened. Before using it on a larger set of records, the team can test it on cases that a person has already reviewed. The comparison helps show which problems it finds, which acceptable cases it flags, and which discrepancies it misses.

That check should be repeated when the data sources, task, or assistant change. Recording the tool and version, the instructions used, and the verification steps makes it easier to explain how the review was performed and to repeat it later.

The records themselves may introduce another risk. A ticket comment or repository file could contain instructions intended to mislead the assistant reading it. This is a form of prompt injection, addressed in OWASP’s guidance. Limiting access, defining the task clearly, and checking the original evidence help reduce the risk of relying on an incorrect response.

The IIA’s AI Auditing Framework addresses governance, management, and internal audit responsibilities. NIST’s AI Risk Management Framework organizes risk work around Govern, Map, Measure, and Manage. These frameworks can help structure an assessment of the assistant’s use. The audit team still has to evaluate its performance on the actual task and verify the answers used in the review.

AI may also appear earlier in the process if developers use it to suggest fixes. GitHub Copilot Autofix, for example, can propose changes for certain code-scanning alerts. That use is separate from the dependency update automation in the example. The auditor can review who checked the proposed code, whether the required tests and security scans ran, and who could dismiss a finding or approve deployment.

I see the main benefit here in reducing the time spent searching through notes and finding related explanations. That can leave more time to examine difficult cases, clarify the policy with the responsible team, and confirm what happened to the affected production version. The audit conclusion still needs to be supported by that review.

References

  1. The IIA. Solving the Riddle: Harnessing Generative AI for Internal Audit Activities.
  2. The IIA. Artificial Intelligence Auditing Framework, 2nd Edition.
  3. The IIA. Global Internal Audit Standards, Standard 14.1.
  4. NIST. Artificial Intelligence Risk Management Framework.
  5. Renovate. How Renovate Works; GitHub. Dependabot Version Updates.
  6. GitHub. About Autofix for Code Scanning.
  7. OWASP GenAI Security Project. LLM01:2025 Prompt Injection.