Google: half of the vulnerabilities found by AI lead to remote code execution

The Google Threat Intelligence Group has observed a doubling in the number of CVEs published each month in 2026, and a vulnerability discovered by the Hacktron agent that was exploited four days after its disclosure.

The Google Threat Intelligence Group (GTIG) has reviewed vulnerabilities published between January 2025 and August 2026. The first observation: the pace is picking up. The number of CVEs disclosed per month has increased from 5,045 in January to 10,740 in August 2026. Only a tiny share (0.23%) is actually exploited, but in volume the increase is clear: 141 vulnerabilities exploited in the first eight months of the year compared to 127 for the whole of 2025, and about 11 zero-days per month instead of 8.

More dangerous AI vulnerabilities

The vulnerabilities attributed to an AI discovery are unlike any other: 50% of them allow remote code execution, compared to 26% for those found otherwise, and their risk profile leans more towards the "medium" level. Case in point: CVE-2026-1731, an authentication-free command injection into BeyondTrust software, spotted by Hacktron's research officer, was exploited four days after its release to elevate privileges, exfiltrate data, and install cryptocurrency miners.

AI software itself is becoming a target: 2,076 vulnerabilities concern them over the period, including more than 1,500 published in 2026. Half of them target agent orchestration frameworks such as Flowise or Langflow.