Rising CVE Volume Puts Exposure Validation at the Centre of Security
AI is accelerating vulnerability discovery while making it harder for security teams to decide which findings warrant immediate action. In the first half of 2026, 35,853 CVEs were published, roughly 49% more than in the previous year. Only 495 were catalogued as exploited in the wild during that period, while 116 were under attack on the day they became public.
Picus Security frames the gap as an exposure-validation problem rather than a simple shortage of vulnerability data. Anthropic disclosure data cited in the article reports that Mythos-class models identified 26,153 vulnerability candidates in open-source software, but only 421 were patched upstream.
Severity is not an environmental verdict
A CVSS rating supplies a shared severity baseline, but it does not establish the likely effect of a flaw in a particular organization. One CVE can occur across hundreds of assets with sharply different risk profiles: some may be unreachable, some protected by controls that disrupt exploitation techniques, and others located on business-critical systems where prevention and detection do not succeed.
The relevant decision therefore depends on evidence from the organization’s own environment. Teams need to establish whether an exposure is exploitable, which assets are affected, whether they can be reached, and how important they are to the business. Treating every High or Critical rating as an emergency becomes impractical as the number of findings rises and the time between disclosure and exploitation narrows.
Why live testing cannot cover every case
Automated and agentic pentesting can provide direct evidence by running real exploits, linking vulnerabilities with credentials and misconfigurations, and demonstrating potential attack paths. However, Picus notes that a usable exploit must exist for CVE-based testing and that the target must be safe to test.
Newly disclosed vulnerabilities may have no working exploit, while live exploitation may be unsuitable for business-critical, restricted or air-gapped systems. Omdia research cited by Picus found that 95% of organizations consider pentesting a top or high priority, but only 32% of their average attack surface is tested each year. These constraints leave exposures that still require an exploitability assessment without a safe live test.
Three forms of validation
Picus distinguishes exploitability validation, security-control validation and agentic pentesting. Exploitability validation assesses whether an exposure can be exploited in the local environment, including cases without a working exploit. Security-control validation tests whether preventive and detective controls block, detect or miss an attack. Agentic pentesting executes exploits and chains exposures to show how far an attacker could progress.
The company argues that these methods should feed one decision process, not operate as isolated workflows. A finding can trigger the validation method suited to its condition, new evidence can alter remediation priority, and a completed fix can be re-validated. For security leaders, the practical implication is to prioritize remediation using validated evidence about exploitability, control effectiveness and asset significance rather than CVSS severity alone.