Hackers target misconfigured proxies to access paid LLM services

Threat actors systematically scanned and probed misconfigured proxy servers and commercial LLM endpoints, generating over 80,000 sessions across 73+ models and leveraging SSRF and model-pull injection attempts; GreyNoise captured two campaigns (one likely using OAST research infrastructure) and recommended mitigations such as restricting model pulls to trusted registries, egress filtering, DNS blocking of known callback domains, and rate-limiting.