AI agents face new security threats from cyber campaigns and platform flaws

New research highlights growing vulnerabilities in artificial intelligence agent frameworks. CrowdStrike has identified four previously unreported threat actors using agentic AI to conduct automated cyber operations. One Spanish/English-speaking actor utilized a dual-agent architecture to target healthcare entities in Argentina, employing a “Supervisor” agent to select tasks and an “Executor” agent to perform them. Another Chinese-speaking operator targeted government and private sector entities globally, including in Brazil and the United States.

Separately, researchers at Black Hat 2026 identified a structural flaw named “CoreBreak” affecting major AI agent platforms from AWS, Google, and Vercel. The flaw demonstrates that traditional guardrails, such as system prompts and content filters, can be bypassed if an attacker bypasses the model itself. Exploits included feeding message histories with pre-existing tool calls or forging “approved” confirmation events to trick human-in-the-loop mechanisms. In response, vendors have issued patches to implement out-of-band policy enforcement, which binds tool execution to states that attackers cannot author.