TLDR Tech / InfoSec
How Attackers Exploited LMDeploy LLM Inference Engines In 12 Hours
Sysdig observed exploitation attempts in its honeypot environment within 12 and a half hours of the publication of an SSRF vulnerability in the LMDeploy vision and text LLM serving toolkit. No public proof-of-concept exploit existed at the time, but the vulnerability disclosure was detailed enough that an LLM was capable of weaponizing it. Sysdig warns that this is becoming the standard and that defenders will need to assume that any vulnerability that is published with substantial information will be exploited nearly instantly.
Read the full story