Hackers bypass AI guardrails using multi-model routing

CrowdStrike researchers have identified four distinct threat groups using agentic AI to conduct autonomous cyberattacks between May and August. These actors, operating in Chinese, Spanish, Indonesian, and Russian, successfully compromised government and private sector targets by routing tasks through multiple LLMs, including older versions of Anthropic's Claude and open-weight Deepseek models. The attackers circumvented safety guardrails by framing malicious activities as authorized security testing or by using runtime patches to remove restrictive system instructions. While the most sophisticated groups built custom 'harnesses' with supervisor-executor agent architectures, even less technical actors achieved successful breaches by leveraging publicly available tools at scale.