AUTO-UPDATED

AI Guardrails Aren’t Stopping Criminal Hackers

CrowdStrike researchers identified four threat groups using agentic AI to automate cyberattacks, successfully bypassing model safeguards to compromise government and private sector targets across multiple global regions.

Key Points

  • CrowdStrike analyzed four threat groups operating in Chinese, Spanish/English, Indonesian, and Russian that utilized AI to conduct autonomous, large-scale cyber campaigns between May and August.
  • Attackers bypassed AI guardrails by using "jailbreaking" prompts, framing malicious activity as authorized penetration testing, and employing runtime patches like ClawGod to remove safety instructions.
  • The campaigns leveraged various models, including Anthropic’s Claude and open-weight Deepseek versions, to perform reconnaissance and exploit known security vulnerabilities.
  • A Spanish and English-speaking group successfully compromised at least 15 Argentine healthcare organizations, stealing sensitive credentials and medical records.
  • The Chinese-speaking actor targeted government and private entities in Brazil, Colombia, and the United States across sectors including biotech, retail, and technology.

Why it Matters

These findings demonstrate that cybercriminals are increasingly using AI to scale operations and exploit basic security hygiene issues that organizations have failed to address. As threat actors refine their ability to bypass model safeguards, businesses must prioritize robust security protocols to defend against automated, agentic attacks.
Wired Published by Lily Hay Newman
Read original