Google Warns AI Reshaping Vulnerability Discovery and Exploitation

Artificial intelligence (AI) is fundamentally altering the landscape of vulnerability discovery and exploitation, according to a new report from Google’s Threat Intelligence Group (GTIG). The analysis indicates a significant increase in both the pace of vulnerability identification and the severity of the flaws being found and actively exploited.

GTIG’s findings, spanning from January 2025 through August 2026, show a doubling in the number of monthly vulnerability disclosures. In January 2026, 5,045 vulnerabilities were disclosed, rising to 10,477 in July and peaking at 10,740 in August. While acknowledging that automated CVE assignment in open-source projects can inflate these numbers, GTIG’s own ratings for high-risk disclosures still grew by 167% during the same period.

The report highlights a measurable shift in the risk profile of AI-discovered vulnerabilities. Fifty per cent of flaws identified by AI agents enable remote code execution (RCE), compared to just 26% of non-AI discovered vulnerabilities. This is attributed to AI models’ enhanced ability to detect memory corruption and logic flaws that often elude traditional static analysis tools.

Rising Exploitation of N-Days

The rate of exploited vulnerabilities has also escalated. GTIG recorded 141 distinct exploited vulnerabilities in the first eight months of 2026, surpassing the 127 observed throughout all of 2025. This equates to an average of 18 exploited vulnerabilities per month, up from 10.5 in the previous year.

While zero-day exploitation saw only a marginal increase, the growth in overall exploitation primarily stemmed from n-days (previously disclosed vulnerabilities). GTIG suggests that threat actors are leveraging Large Language Models (LLMs) and other AI tools to automate the analysis of product version differences, patches, and proof-of-concept (PoC) code, enabling them to rapidly weaponise n-days rather than investing time in discovering new zero-days.

In-the-wild exploitation of AI-discovered vulnerabilities has also been confirmed. One notable example is CVE-2026-1731, an unauthenticated OS command injection flaw in BeyondTrust Privileged Remote Access and Remote Support. This vulnerability was autonomously discovered by the Hacktron AI research agent and subsequently exploited by a threat cluster within four days of public disclosure, with five more clusters following within a week.

Vulnerabilities in AI Systems Themselves

Beyond AI’s role in discovering traditional software flaws, GTIG also tracked a rise in vulnerabilities affecting AI systems themselves. Between January 2025 and August 2026, 2,076 AI-related CVEs were identified, with over 1,500 appearing in 2026 alone. Approximately half of these affect AI orchestration frameworks.

While only a handful of these AI infrastructure vulnerabilities, such as flaws in LiteLLM and Langflow, have been confirmed as exploited in the wild, GTIG has not yet observed zero-day exploitation targeting AI infrastructure. The group anticipates that both vulnerability discovery and exploitation rates are likely to continue their upward trend in the short to medium term.