From WIRED · Lily Hay Newman · · 2m
Criminal hackers are using AI to bypass safety guardrails
Four threat groups have successfully weaponised AI models by circumventing their safeguards, targeting government and private organisations across multiple countries.
In brief
Four threat groups have successfully weaponised AI models by circumventing their safeguards, targeting government and private organisations across multiple countries. CrowdStrike researchers found four threat groups using AI models to conduct autonomous hacking campaigns between May and August, bypassing model…
Four threat groups weaponise AI
CrowdStrike researchers identified four previously unreported threat groups using agentic AI in attacks between May and August. The groups operated in Chinese, Spanish and English, Indonesian, and Russian.
Models used and infrastructure exposed
The attackers primarily used versions of Anthropic's Claude and open-weight Deepseek models. All four groups left their AI infrastructure openly accessible on web servers, allowing researchers to analyse their activity.
Older models, lower costs
Attackers used older model versions rather than frontier models to reduce costs. Adam Meyers, CrowdStrike's senior vice president of counter adversary work, notes this mirrors legitimate AI usage patterns.
“You can leverage different models for different aspects in the operation. Model routing and selection is very important because it's expensive and a lot of times would be overkill to use newer models.”
Exploiting guardrail differences
The attackers used multiple models to take advantage of differences in safeguards between them. Models do not all have equivalent guardrails and protections, even from one company.
“The models don't all have equivalent guardrails and protections, even from one company.”
Jailbreaking methods and deception
All four threat actors circumvented AI safeguards by convincing models their activity was legitimate, often framing attacks as authorised penetration testing. They added files to working directories to show approval and used traditional jailbreaking prompts.
Spanish-speaking group targets healthcare
The Spanish and English-speaking actor targeted Argentine health care entities and successfully compromised at least fifteen organisations. Researchers found evidence the actor stole credentials and medical records from at least one victim.
Chinese-speaking group's wider reach
The Chinese-speaking actor successfully compromised government targets in Brazil and Colombia and targeted private organisations in the United States and Brazil across biotech, healthcare, food and beverage, retail, technology, and travel sectors.
Less sophisticated groups still succeed
The Russian-speaking and Indonesian-speaking actors used more basic, publicly available skills but still successfully compromised targets including multiple Indonesian government domains, a Bangladeshi government domain, and private-sector targets.
The key point
CrowdStrike researchers found four threat groups using AI models to conduct autonomous hacking campaigns between May and August, bypassing model safeguards by posing attacks as authorised security testing.
