Cybersecurity firm CrowdStrike reported Wednesday that four distinct threat groups have successfully deployed agentic AI systems to automate cyberattacks between May and August. The actors operated in Chinese, Spanish/English, Indonesian, and Russian, utilizing both open-weight models like DeepSeek and closed-weight models like Anthropic’s Claude to execute autonomous campaigns over extended periods.

According to CrowdStrike, the attackers regularly bypassed model safety guardrails using social engineering prompts, such as framing malicious activity as legitimate penetration testing. Sophisticated actors employed custom runtime patches, including a tool named ClawGod designed to alter Claude Code system prompts, while less sophisticated groups relied on basic jailbreaking techniques to force model compliance.

Researchers noted that these campaigns largely exploited older AI models to control operational costs while targeting routine security hygiene vulnerabilities. The findings indicate that threat actors are leveraging agentic tools to increase the volume and velocity of conventional cyberattacks rather than relying on novel exploits.

Why it matters

  • Demonstrates that current AI guardrails can be consistently bypassed by adversaries using basic jailbreaks and system prompt modifications.

  • Shows how threat actors use model routing and older open/closed models to optimize the cost and scale of automated attack infrastructure.

  • Signals urgent requirements for enterprise security teams to defend against high-frequency, AI-driven automation of known vulnerabilities.

Source: wired.com