Getty Images

News

AI-discovered flaws are rarely exploited but pose greater risk

Relatively few AI-discovered flaws are exploited, but new research shows the ones that are can pose outsized risk to enterprises.

Alissa Irei
By
Published: 13 Oct 2026

Attackers are exploiting only a tiny fraction of AI-discovered vulnerabilities -- but recent analysis suggests the ones they do exploit tend to be more dangerous to enterprises.

As previously reported by TechTarget, only about 1% of vulnerabilities attributed to AI-assisted discovery were listed as exploited in the wild as of June 2026. That's according to Patrick Garrity, a security researcher at vulnerability prioritization tool maker Vulncheck, who analyzed data from Anthropic's vulnerability disclosure dashboard, the Berkeley Vulnerability Research Initiative and Vulncheck's own known-exploited vulnerability data.

Among publicly disclosed vulnerabilities, however, those found by AI tend to have more serious exploitation consequences than conventional CVEs, according to a separate report from the Google Threat Intelligence Group (GTIG). Half of all AI-discovered vulnerabilities resulted in remote code execution (RCE), GTIG found, compared to just 26% of CVEs discovered by humans or conventional scanners. GTIG also rated 69% of conventional CVEs as low risk, while 60% of CVEs discovered by AI were medium or high risk.

GTIG based its research on confirmed disclosures from verified lab and vendor ledgers, CISA advisories, MITRE records and vendor security bulletins.

For security leaders, the implication is that exploit frequency might be low, but the potential enterprise impact of the exploited subset could be higher than traditional vulnerability patterns would suggest.

The pattern is not surprising, according to Kelli Vanderlee, senior analyst at GTIG and co-author of the report. That's because threat researchers deliberately prompt autonomous agents to audit critical infrastructure and sensitive privilege boundaries, with a focus on high-impact findings. The agents, in turn, tirelessly pursue novel ways to meet their objectives with a single-minded persistence and technical fluency that humans and conventional scanners lack.

"Autonomous systems are engineered to navigate complex, multi-step semantic code paths across core C/C++ libraries, runtimes and hypervisors," Vanderlee told TechTarget. "By synthesizing fuzzing harnesses, modeling memory states and chaining obscure edge-case logic, AI models excel at identifying memory corruption -- buffer overflows, use-after-free -- and logic bypasses that consistently elude traditional static analyzers."

Real-world risk of AI-discovered flaws

In a real-world example of AI-driven RCE risk, a research AI agent at Hacktron AI autonomously discovered CVE-2026-1731, an unauthenticated OS command injection flaw in BeyondTrust Privileged Remote Access and Remote Support. Within four days of public disclosure, GTIG researchers observed a threat cluster exploiting the vulnerability in targeted initial-access campaigns, with five more threat clusters appearing within the first week. An array of post-exploitation activities followed, including data exfiltration, privilege escalation and secondary payload drops.

GTIG experts called their RCE findings an early indicator rather than an established trend. They added, however, that the BeyondTrust example highlights real-world implications, with threat actors actively exploiting high-impact vulnerabilities that defensive AI agents find.

A caveat: Context still determines risk

It's unsurprising that exploited AI-discovered vulnerabilities tend to disproportionately enable RCE, said Matthew Smith, a virtual CISO and management consultant specializing in cybersecurity risk management and AI. That's because LLMs and AI agents are most effective when given an objective function.

"In this case, 'maximize remote code execution' is a pretty clear goal that an attacker or researcher could state when looking for vulnerabilities," Smith said.

But a flaw's relative severity depends heavily on context, he cautioned. GTIG researchers, Smith noted, use Google's own vulnerability-scoring methodology, rather than the Common Vulnerability Scoring System. Google says it prefers a sharper risk-based severity triage that results in fewer "critical" ratings.

"What that tells me is that even reasonable people can disagree on what 'critical risk' is," Smith said. "Therefore, how those vulnerabilities are being categorized by a given system may or may not apply to a given organization's actual deployment."

Ultimately, organizations must independently assess and triage vulnerabilities based on risk -- whether AI found them or not.

Alissa Irei is an Informa TechTarget news reporter covering cybersecurity.

Dig Deeper on Threats, Cyberattacks & Vulnerabilities