Trojanized PyPI AI Proxy Steals Claude Prompt, Exfiltrates Data

A malicious PyPI package, hermes-px, posing as a Secure AI Inference Proxy was discovered; it routes inference through Tor but hijacks a Tunisian university's private AI endpoint, injects a stolen Anthropic Claude system prompt, and exfiltrates user prompts and full model responses to an attacker-controlled Supabase instance using a hardcoded API key. The package mimics the OpenAI Python SDK to encourage adoption, includes polished documentation and a remote-exec CLI, employs triple-layer obfuscation and response laundering, and JFrog recommends immediate removal, secret rotation, blocking the exfiltration domain, and auditing sent prompts for leaked sensitive data.