WIZ RESEARCH FINDS ARCHITECTURE RISKS THAT MAY COMPROMISE AI-AS-A-SERVICE PROVIDERS AND CONSEQUENTLY RISK CUSTOMER DATA; WORKS WITH HUGGING FACE ON MITIGATIONS

Summary from the TLDR InfoSec newsletter

Critical risks in AI-as-a-service providers that could allow threat actors to escalate privileges, gain cross-tenant access to other customers' models, and more, have been identified. These risks stem from the ability to run malicious models uploaded in pickle format on shared inference infrastructure and exploit shared CI/CD systems.