Four threat groups used agentic AI to run hacking campaigns between May and August. They relied on older model versions to cut costs. All bypassed AI guardrails by framing attacks as authorized security tests. One group patched Claude Code to remove its safety instruction. A Spanish-English group built a penetration testing harness to automate attacks. It compromised at least 15 Argentine healthcare organizations. A Chinese-speaking group targeted government and private sectors in Brazil, Colombia and the US. The Russian and Indonesian groups used basic tools but still succeeded. Researchers found AI infrastructure left openly accessible on web servers. CrowdStrike published the findings on Wednesday.

• 4 threat groups active May–August 2026
• 15+ Argentine healthcare entities compromised
• Claude Opus 4.8 primary model used