Summary
- CrowdStrike researchers discovered four previously unreported threat groups using agentic AI and bypassing model safeguards between May and August.
- The groups, operating in Chinese, Spanish/English, Indonesian, and Russian, leveraged open- and closed-weight models such as Anthropic Claude and DeepSeek.
- Attackers used older model versions, model routing, and jailbreaking techniques like framing malicious tasks as authorized penetration testing.
- The Spanish/English and Chinese groups employed sophisticated autonomous harnesses and runtime patches to orchestrate attacks against health care, government, and private targets.
Why it matters
Cybercriminals are successfully utilizing agentic AI and bypassing model guardrails to execute autonomous, large-scale attacks using standard security vulnerabilities. This demonstrates that current AI safety measures and model protections are insufficient to prevent malicious exploitation.