The vulnerabilities AI finds are the ones attackers want

Google Threat Intelligence Group studied vulnerability disclosures and exploitation from January 2025 through August 2026 and found that flaws discovered by AI research agents are weaponized unusually fast. A standout case is CVE-2026-1731, an unauthenticated OS command injection bug in BeyondTrust Privileged Remote Access and Remote Support identified by the Hacktron AI agent: one threat cluster began exploiting it within four days of public disclosure, and five more clusters followed within a week, deploying SNOWLIGHT, SPARKRAT, and cryptominers after escalating privileges and stealing data.

GTIG also reports that half of the vulnerabilities it attributes to AI discovery result in remote code execution, versus just 26% of those found through other methods, and that overall exploited CVE counts in 2026 already surpass all of 2025. AI infrastructure itself is increasingly targeted: agent orchestration tools like Flowise and Langflow account for half of this year's AI-related disclosures, while inference platforms including vLLM, Ollama, and LiteLLM saw actively exploited bugs such as CVE-2026-42271, CVE-2026-5027, and CVE-2025-3248, enabling host takeover, credential theft, and remote code execution.