Rising Threat: AI-Driven Exploitation of Publicly Disclosed Vulnerabilities

Recent analyses indicate a disturbing trend in cybersecurity: attackers are increasingly leveraging publicly disclosed flaws instead of seeking new zero-day vulnerabilities. This shift appears to be accelerated by AI tools, which streamline the process of weaponizing known weaknesses.

A report from Google’s Threat Intelligence Group (GTIG) reveals that between January and August 2026, there were 141 exploits in the wild, surpassing the 127 recorded for all of 2025. This marks a significant uptick, coinciding with an explosive increase in vulnerability disclosures—growing from 5,045 in January to nearly 10,740 by August. In terms of exploitation frequency, incidents went from an average of about 10.5 per month last year to close to 18 in 2026. Notably, while zero-day exploits saw a slight rise from eight to eleven monthly, they still pale in comparison to the rapid growth of n-day vulnerabilities.

Security professionals are becoming increasingly concerned about n-day vulnerabilities, as evidenced by findings from Foundry’s upcoming Security Priorities survey. The results indicate that 38% of respondents are deeply worried about software n-day exploits, while 37% express similar anxiety about network-based n-day threats—contrasting the 25% who view zero-day exploits as their main concern.

According to GTIG's analysis, the efficiency of threat actors utilizing large language models (LLMs) and AI tools for their exploits is noteworthy. These technologies allow attackers to automate the analysis of product versions, patches, and vulnerability notifications—facilitating rapid exploitation of known flaws, rather than the slower burden of discovering new vulnerabilities.

Doubling Down on Disclosures

The escalation in new vulnerabilities does not occur in isolation. GTIG cautions that some of this influx may be attributed to automated CVE Numbering Authority policies within open-source ecosystems, effectively inflating baseline figures. For instance, vulnerabilities specifically related to the Linux kernel alone accounted for nearly 5,000 CVEs this year, even though no zero-days were seen exploited during this period.

Some vendors have also recorded significant spikes in high-risk flaws. Totolink, for example, dealt with 75 high-risk vulnerabilities in its consumer router firmware during just two months. Oracle experienced a similar burst following its quarterly Critical Patch Update, leading to 128 high-risk vulnerabilities reported in August.

Despite the increase in disclosed flaws, the actual percentage being exploited remains deceptively small at roughly 0.23%, or about one in every 431 vulnerabilities disclosed. However, there's a clear correlation between the growth of vulnerability disclosures and exploitation indices, suggesting that while exploitation might not currently overtake disclosure rates, the gap is narrowing.

On a related note, analytics from VulnCheck observed a reduction in the proportion of flaws being exploited as vulnerability disclosures climbed—falling to approximately 1.4% in H1 2026, following a peak of 2.7% in 2023. Nevertheless, the median time from CVE publication to exploitation has plummeted from 120 days in 2025 to 80 days in early 2026, likely highlighting the role that AI agents are playing in expediting these attacks.

Targeting Vulnerable Perimeter Defenses

A significant number of vulnerabilities exploited from January to August 2026 are associated with edge and security appliances, which account for around 14% of incidents. Notably, a majority of these vulnerabilities are categorized as high or critical severity. This trend underscores a growing focus on targeting perimeter defenses, which have frequently been entry points for both state-sponsored actors and ransomware groups.

Attackers tend to exploit unauthenticated management interfaces on public-facing systems, evading detection by enterprise endpoint protection solutions. GTIG researchers emphasize that while CVE discovery is surging, exploitation activity is still predominantly directed toward perimeter devices and openly exposed enterprise services.

The Severity of AI-Discovered Flaws

While discerning which vulnerabilities are detected via AI assistance can be opaque, GTIG found that over half of the vulnerabilities identified by AI models are rated medium severity or higher. In contrast, less than one-third of all disclosed vulnerabilities fall into this category. Alarmingly, 50% of AI-discovered flaws involve remote code execution as a potential impact, compared to just 26% among traditional disclosures.

The GTIG report suggests that this pattern reflects the methodology behind AI research programs, which often prioritize auditing critical infrastructure rather than scouring for minor flaws. As the narrative unfolds, there is mounting evidence that threat actors are also deploying AI tools to examine software patches and develop exploits for n-day vulnerabilities. For example, a recent GTIG report detailed how a North Korean APT group sent thousands of queries to an AI model to analyze public vulnerabilities and validate proof-of-concept exploits.

Bidding Farewell to Mass Patch Management

With both vulnerability discovery and exploitation rates on the rise, organizations need to rethink their approaches. Relying solely on mass patching is no longer viable. As highlighted by Detectify, a surveillance firm that analyzes external attack surfaces, critical and high-severity flaws in exposed assets often remain unaddressed for over 90 days. This includes organizations experiencing a marked increase in their verified exposed domains—20% in the US and 14% in the UK in just a year.

On September 22, the Cybersecurity and Infrastructure Security Agency (CISA) unveiled a framework aimed at transitioning the CVE Program from its current growth phase into a quality-centric era. This comes in response to over 67,000 recorded CVEs through mid-September, with projections suggesting that number could reach upwards of 96,000 by year’s end. However, experts like Farzad Bakhtiar of Flashpoint caution that merely having well-defined records doesn’t equate to actionable insights.

“Security teams need context on vulnerabilities—specifically whether they have working exploits, if they’re being actively used, and if they touch any exposed assets,” Bakhtiar emphasizes. With the CVE count set to near 100,000 this year and many vulnerabilities foregone altogether, effective contextualization will be essential for transforming raw vulnerability data into prioritization decisions.