Forecasting the Cybersecurity Vulnerabilities of Agentic AI Systems toward 2030: A Framework for Autonomous AI Security

Ofuka Abung Abung *

University of Fairfax, 1813 E Main St, Salem, VA 24153, United States of America.

Oluwaseun Oladeji Olaniyi

University of the Cumberlands, 104 Maple Drive, Williamsburg, KY 40769, United States of America.

Fidelia I. Ahiante

Prairie View A&M University, 100 University Drive, Prairie View, Texas 77446, United States of America.

Omobolaji Olufunmilayo Olateju

University of Ibadan, Oduduwa Road, Ibadan, Oyo State, Nigeria.

Oluwadayo Mafolasere Olaniyi

University of the Cumberlands, 6178 College Station Drive, Williamsburg, KY 40769, United States of America.

*Author to whom correspondence should be addressed.


Abstract

Aims: Agentic artificial intelligence (AI) systems plan, call tools, hold credentials, and act with little human supervision, yet their security is still assessed mostly after incidents occur. This study forecasts vulnerability disclosures in a selected set of open-source agentic AI components to 2030 and proposes a security framework that responds to those forecasts.

Methodology: Four public datasets were assembled: (i) 943 unique, dated vulnerability disclosures (April 2023 to October 2026) in 87 open-source agentic components from PyPI and npm, enriched with CVE records, CVSS vectors, and CISA SSVC decisions; (ii) the CISA Known Exploited Vulnerabilities (KEV) catalogue; (iii) a census of 30,054 Model Context Protocol (MCP) and agent projects on PyPI; and (iv) MITRE ATLAS. Six forecasting models were backtested from four origins, and the identifiable models were combined in a quantile-averaged ensemble.

Results: Annual disclosures rose from 19 in 2023 to 164 in 2025 and reached 705 by mid-October 2026, with an April 2026 changepoint that lifted the monthly mean from 8.9 to 94.5. Agent platforms contributed 54% of disclosures, and access-control flaws overtook code injection as the largest weakness family. Nine agentic platform flaws entered KEV in 2026, with a median interval of 30 days between publication and listing; because KEV listing follows confirmed exploitation, this interval is an upper bound on the time to first exploitation. Saturating models were not identifiable because discovery has not reached its inflection point. The ensemble projects a median of 5,334 disclosures in 2030, but the 80% interval (495 to 53,088) is very wide, so this figure indicates a planning range rather than a point prediction. This estimate is bracketed by a diffusion-bounded plateau near 1,200 disclosures per year and an acceleration envelope.

Conclusion: We propose the Anticipatory Autonomous AI Security (A2AS) framework, which nests runtime, platform, forecast, and governance controls around the agent and escalates control posture when observed disclosures exceed forecast bands. In a limited retrospective test from two forecast origins, it would have escalated two months before the 2026 surge from one origin and in the month of the surge from the other. A2AS is therefore offered as a proposed framework whose alert thresholds require prospective testing.

Keywords: Agentic AI, vulnerability forecasting, model context protocol, autonomous AI security, prompt injection, known exploited vulnerabilities, probabilistic forecasting


How to Cite

Abung, Ofuka Abung, Oluwaseun Oladeji Olaniyi, Fidelia I. Ahiante, Omobolaji Olufunmilayo Olateju, and Oluwadayo Mafolasere Olaniyi. 2026. “Forecasting the Cybersecurity Vulnerabilities of Agentic AI Systems Toward 2030: A Framework for Autonomous AI Security”. Journal of Engineering Research and Reports 28 (11):181-97. https://doi.org/10.9734/jerr/2026/v28i112051.

Downloads

Download data is not yet available.