AI Squeezes the Window for Vulnerability Remediation

CVE volume reached 48,000 last year and is projected to exceed 100,000 this year, according to Forrester senior analyst Janet Worthington. That would more than double last year's total.

Worthington made the forecast at this week’s “Mythos: Breaking Out of the Vulnerability Matrix” session during Forrester’s Security & Risk Forum in Washington, D.C. The Common Vulnerabilities and Exposures list comes from the Department of Homeland Security. AI models can discover flaws faster, generate potential exploits and help developers remediate code that once required extensive human analysis.

Historically, teams had time between disclosure and exploitation to assess risk, prioritize remediation and deploy patches. AI is compressing that window, and conference speakers said some vulnerabilities can now be exploited before public disclosure, leaving fewer chances to respond.

From vulnerability counts to exposure

Worthington said AI changes not only how security programs are set up and how teams respond, but also the capabilities available to attackers. Counting vulnerabilities no longer tells leaders enough.

“In the past … we spent a lot of time counting vulnerabilities, being very proud that we could give those numbers to the executives. Because AI is accelerating, both discovery and exploitation, we really need to talk in terms of exposure now,” she said.

Exposure means establishing which systems and data are at risk, whether existing controls prevent exploitation and whether remediation meaningfully reduced risk. Security teams need continuous exposure validation, not just periodic scans or traditional severity ratings, and they must measure effectiveness rather than celebrate identified or patched totals.

That shift requires leaders to rethink risk validation, remediation priorities, AI deployment and effectiveness measurement while preserving human oversight. Erik Nost, another Forrester senior analyst, cautioned that a CVE is not automatically a business risk: it may be unreachable or unexploitable in a given environment. He urged teams to validate vulnerabilities in their own infrastructure rather than infer exposure from a list.

AI has not earned trust to patch code

AI gives application-security teams capabilities they previously could not easily scale. Models can analyze large codebases, identify complex vulnerabilities and propose fixes. Historically, automation worked best for straightforward changes such as updating vulnerable third-party libraries; proprietary code demanded understanding of application behavior, architecture and business context.

AI can supply that context and generate code changes, but security teams should not assume those changes are safe to deploy without review. Agents can recommend fixes and automate parts of the workflow, including ticket creation and suggested changes, while deployment still requires knowledge of application behavior, infrastructure, dependencies, business priorities and operational consequences.

Nost was blunt: “We don't trust AI [for remediation],” he said, citing studies in which different models ran against 600 vulnerabilities. “We found that AI only fixed the vulnerability without breaking the application and without introducing a new vulnerability only 26% of the time.”

A seemingly simple change can disrupt customer access, interrupt revenue-generating services or introduce a new flaw. Forrester speakers set four conditions for AI-assisted remediation:

• Complexity: Novel vulnerabilities and unfamiliar environments warrant additional human review.
• Business impact: Changes affecting availability, customers or revenue should involve human decision-makers.
• Accountability: AI-generated code should not replace established development ownership and responsibility.
• Context: Teams need evidence of potential consequences, including exposure scope and the blast radius of proposed changes.

Over time, Forrester expects more autonomous remediation for well-understood, low-risk changes, paired with approvals and rollback. People should remain the default for consequential decisions, even as agents handle repeatable work.

Three priorities for security leaders

Forrester analysts outlined three capabilities that organizations should prioritize as AI reshapes security operations:

• AI red teaming: Businesses developing AI agents or engineering systems should test them against adversarial behavior. Human-led testing remains necessary to identify unexpected ways attackers might manipulate models, agents and associated tools.
• Exposure validation: Teams must establish whether vulnerabilities are reachable and exploitable in their environments, separating theoretical weaknesses from actual business risk.
• Agentic development security: As developers adopt coding agents, teams need controls for insecure code generation, vulnerable dependencies, exposed secrets and malicious instructions. They also need to address risks in Model Context Protocol integrations and other agent extensions.

The speakers also urged teams to prioritize critical applications, remove unnecessary attack surface, strengthen software-development pipelines and invest in developer security champions who can bring security expertise into engineering teams.

What does this change for security teams?

For enterprises, the operating model must shift from periodic, severity-led scanning toward continuous validation of real exposure. AI can accelerate discovery, code analysis and portions of remediation, but leaders must pair it with contextual risk decisions, established development ownership and human approval for consequential changes.

AI does not remove established security fundamentals; it increases the speed and scale at which teams must apply them. AI can accelerate the patch cycle; it cannot, on this evidence, replace the judgment required to keep applications intact.