# Agentic AI Risks and the CVE Program at Black Hat 2026

> Black Hat USA 2026 spotlighted two big issues: the security risks of agentic AI and growing doubt over the effectiveness of the CVE program.

**URL:** https://www.ciptadusa.com/blog/agentic-ai-cve-program-20260828  
**Type:** blog  
**Author:** PT Cipta Dua Saudara  
**Category:** Application Security  
**Published:** 2026-08-28  
**Cover:** https://cdn-uagents.enitip.com/uploads/blog/2026-08/daily-appsec-20260828-072349.jpg  

## Article

At Black Hat USA 2026, two concerns dominated security experts' conversations at once: the surging risk of agentic AI, and questions about the future of the CVE program. The two are not a coincidence. As AI agents begin executing tasks inside production systems, the attack surface grows as fast as mitigation catches up. And amid that shift, the vulnerability-classification framework that has anchored global security for two decades faces a credibility crisis of its own.

## Summary

Black Hat USA 2026 spotlighted two big issues: the security risks of agentic AI and growing doubt over the effectiveness of the CVE program.

## Background

Agentic AI differs from ordinary AI models. Instead of merely answering a question, an agent is designed to take action — run commands, call APIs, move data. When that authority lands inside production systems without adequate oversight, a single manipulated instruction can turn an agent from an assistant into an entry point for attack. Researchers at Black Hat framed this as a new risk class that lacks mature mitigation standards.

Side by side, the CVE program that has underpinned the security ecosystem for years is meeting skepticism. Many researchers question the quality and consistency of CVE entries, especially as numerical standards like severity scores stop reflecting real risk in AI-based architectures.

## The Challenge

The first challenge is **authority and isolation**. Giving an agent the right to act means creating an opening if that right is abused or manipulated. No granular permission control is yet sufficient to guarantee an agent touches only what it is allowed to.

The second challenge is **traceability**. In a traditional system, an exploitation can be traced to specific logs and commands. In a system with agents, decision flows can overlap across automation layers, making it hard to assign responsibility for an incident.

## Approach

The approach researchers recommend revolves around three principles. **Least privilege** — an agent is granted only the minimal access needed for its task, with every action monitored separately. **Human approval** — high-risk actions such as data transfer or critical configuration changes require human validation, not full automation. **CVE re-engineering** — the vulnerability-classification process needs reform to be more transparent, consistent, and aligned with the realities of new architectures.

For organizations starting to adopt agentic AI, the message is simple: security is not an add-on but a design requirement from the start. The earlier decisions about permissions and oversight are made at the design stage, the smaller the risk that has to be borne after the system goes live.

## References

- [Agentic AI Risks, CVE Program Concerns Permeate Black Hat USA 2026 (Dark Reading)](https://www.darkreading.com/cybersecurity-operations/agentic-ai-risks-cve-program-concerns-black-hat-usa-2026)
- [Agentic AI Security Guidance (OWASP)](https://genai.owasp.org/)
- [CVE Program and CNA updates (CVE.org)](https://www.cve.org/)

---

*Markdown version of https://www.ciptadusa.com/blog/agentic-ai-cve-program-20260828 — generated for AI agents and LLM crawlers.*