A security team rolls out an AI scanning tool and receives a backlog of hundreds of issues the next day. Management may see a major improvement in detection capabilities, but the development team faces a series of practical questions: Which vulnerabilities should be addressed first? Who is responsible for making the changes? Is there room in the testing schedule? When can the system be taken offline for deployment? Unless these questions are resolved in advance, the faster AI finds vulnerabilities, the faster the backlog will grow.
TechOrange reported that Bank of America added 1,000 external contractors in the second quarter to work alongside its existing 3,000 internal employees on software vulnerabilities uncovered by AI. The models can identify hundreds or even thousands of low- to medium-risk vulnerabilities. Some issues that might previously have taken weeks to patch now need to be handled within days. The lesson for enterprises is straightforward: Before purchasing an AI security tool, a company should make sure it has enough capacity to fix the vulnerabilities, manage the entire remediation process, and retrieve a record of every action afterward.
Calculate remediation capacity before expanding the scope of scanning
When evaluating tools, many enterprises begin by asking how many vulnerabilities a tool can find and which codebases it can scan. They often fail to account for all the work that follows. Once a vulnerability has been identified, the team still needs to validate the risk, locate the system owner, modify the code, schedule testing, and determine when to deploy the patch. If the problem lies in a vendor package, procurement, legal, or the vendor's support team may also need to become involved. A scan report is only the starting point. Finding a vulnerability does not mean the risk has already been reduced.
Management can begin with a capacity exercise for one category of systems, such as an internal quotation system, a customer service backend, or a factory quality control system. Rather than scanning every codebase from the outset, the company should first record how many vulnerabilities its current team can assess, patch, test, and deploy each week, then identify where work most often gets stuck. The goal at this stage is not to produce an impressive remediation rate. It is to understand which department will be the first to fall behind when the backlog suddenly grows.
The remediation order should not be based solely on the risk score assigned by the tool. The assessment should also account for the importance of the system, whether it is externally accessible, which data it can access, what safeguards are already in place, and how much downtime a failed patch could cause. Security staff and the system owner should make the final decision. If the team cannot process even one category of systems within the agreed timeframe, it should adjust work assignments and deployment windows before expanding the scope of scanning.
Which stages of the patching process require human approval?
AI can help organize vulnerability descriptions, compile relevant code, and prepare work cards. Risk classification, however, still requires human judgment, and code changes should not move automatically into production without review. A practical approach is to divide the process into four stages and document who is responsible for each stage, what information is required, what qualifies as completion, and what happens when a deadline is missed.
Vulnerability classification: AI first compiles the affected systems, the data that may be exposed, and a recommended severity level. Security staff then determine whether the priority should be raised or lowered. If information is missing, the work card must be returned for completion rather than allowing the model to make assumptions and continue.
Remediation assignment: The system owner must decide whether the case will be handled by the internal development team, the software vendor, or an external contractor. The company must also restrict in advance which code and documents the assigned party can access. The work card should state the deadline clearly. If the issue involves a dependent package or a legacy system that cannot be modified, it should also specify who will decide on compensating controls.
Testing and validation: Based on the changes made, testers must verify that the vulnerability has been resolved while confirming that existing functions, such as login, quotations, orders, or production-line integrations, continue to work. The person who made the change must not be allowed to approve their own work, and the test evidence should be attached to the same work card.
Deployment approval: The system manager determines the deployment window based on the risk, test results, and impact of downtime. Changes that carry greater risk or could affect core operations require another approval before the production environment is updated. EgentWrX can turn tasks into reusable work cards, connect multiple tasks into workflows, and require human approval before a handoff so that work proceeds only after the designated person has signed off. An enterprise can begin by documenting the four-stage process for one category of systems, then use real cases to verify that someone is genuinely accountable at every approval gate.
How should enterprises manage vulnerability backlogs as they add contractors?
External contractors can provide additional remediation capacity, but the vulnerability lists they access may expose weaknesses in internal software and vendor code. Research notes indicate that enterprises need to confirm how long contractors are expected to work, how accounts and permissions will be handled when personnel leave, and who will audit their access to live vulnerability data. If a company relies only on nondisclosure agreements without setting account expiration dates and data access boundaries, unnecessary permissions may remain in place after the project ends.
Enterprises can begin by creating a contractor access register. At a minimum, it should record the contractor's employer, the business units and data they can access, the approver, the start date, and the expiration date. Contractors do not need access to the complete vulnerability list and should never share accounts. Security or audit personnel should review access logs each month to determine who viewed, downloaded, or modified data. When a project ends or a contractor leaves, disabling the account, revoking permissions, and reviewing access records should be mandatory closure requirements.
EgentWrX can restrict the actions users may perform and the data they may view according to their organization and role. EgentWrX also maintains audit logs of actions taken by AI Agent and administrators and can verify whether those records have been altered. Platform controls, however, cannot replace clear internal accountability. The enterprise must still designate who approves contractor access to data, who reviews access records regularly, and who is responsible for suspending access and notifying the relevant departments when an anomaly is detected.
Finally, when evaluating AI security tools, enterprises must include the additional costs of remediation, testing, downtime, and contractor management. Bank of America spent more than $200 million in the third quarter on software remediation and anti-money laundering compliance issues combined. The original report did not break down the amount allocated to each, so the figure cannot be treated as the cost of software remediation alone. A more practical approach is to run one category of systems through several complete cycles, record the labor and waiting time at each stage, and then decide whether to expand the scope of scanning. This allows the enterprise to understand how much work its current process can handle before AI uncovers even more vulnerabilities.
FAQ
Does every vulnerability found by AI need to be patched immediately?
Not every vulnerability needs to be fixed at the same time, but each one must enter the classification and assignment process. Enterprises should assess the tool's score alongside the importance of the system, the scope of potential data exposure, existing safeguards, and the impact of downtime. Security staff and the system owner should then confirm the remediation deadline. This prevents low-risk cases from consuming resources and delaying genuinely urgent patches.
Can AI Agent modify code and deploy it automatically?
Allowing AI Agent to modify code and automatically deploy it all the way to production is not recommended. Enterprises can use AI Agent to organize information, prepare work cards, and connect workflows, but designated personnel should still approve the risk assessment, test results, and deployment window. Before updating a core system, the team should also confirm the rollback procedure and downtime arrangements.
What permissions do external contractors need?
Contractors need access only to the code, documents, and testing environments required to handle their assigned cases. The company should define the data access scope, approver, and expiration date for every account, prohibit shared accounts, and disable accounts, revoke permissions, and review access records when a project ends or a contractor leaves.
How can an enterprise determine whether its current team can handle the increase in vulnerabilities generated by AI?
Start with a capacity exercise for one category of systems. Record the labor required for vulnerability classification, remediation, testing, and deployment, as well as how long cases wait at each stage. If cases continue to accumulate at one stage, adjust the responsible personnel, completion criteria, or deployment windows before expanding the scope of scanning. Do not judge capacity solely by how many vulnerabilities the tool can find.
References
- Who Will Fix the Vulnerabilities Found by AI? Bank of America Adds 1,000 Contractors in a Rare Public Disclosure of Its Remediation Workforce — TechOrange reports on the remediation workforce, costs, and contractor-management risks associated with Bank of America's response to vulnerabilities uncovered by AI.
