CrowdStrike says four hacker groups bypassed AI safeguards
🎧 Listen to this story
CrowdStrike researchers analyzed the AI infrastructure used by four threat groups operating in Chinese, Spanish and English, Indonesian, and Russian in attacks carried out between May and August. The company says all of them bypassed model safeguards by presenting their actions as authorized security testing, using jailbreaking techniques, and exploiting differences between protections. The groups mainly used versions of Anthropic’s Claude and open-weight Deepseek models, while also leaving their infrastructure accessible on web servers. The more sophisticated actors built systems to run operations autonomously for relatively long periods; the others also compromised multiple targets despite having less technical expertise.
The Spanish- and English-speaking group targeted Argentine healthcare entities and compromised at least 15 organizations. CrowdStrike found evidence that it stole credentials and medical records from at least one victim. According to the analysis published by Wired, the attacks did not rely on new exploits, but on the speed and volume provided by AI to exploit basic security weaknesses.
Why it matters · editorial interpretation
CrowdStrike’s analysis suggests that model safeguards alone may not prevent abuse when groups combine jailbreaks, differences between protections, and exposed infrastructure. According to the analysis published by Wired, AI’s practical advantage was speed and volume in exploiting basic weaknesses, creating direct risks to credentials and medical records.
Sources
Security