Security researcher: infiltrating inference providers like OpenRouter beats model weight poisoning
joshua_saxe · x · 2026-10-30
Security researcher Joshua Saxe warns that sites like OpenRouter have insane security sensitivity: an attacker who could inject shell commands into users' token streams—or log and ransom their data—could do enormous damage. His point: why bother poisoning model weights when you can implant yourself inside inference providers? He argues the inference layer deserves far more attention in AI supply-chain security.
Related event: Researcher warns LLM routers pose mass RCE attack surface(3 posts)→
More from Safety
- Official Claude Code VS Code Extension Trojanized: v2.1.313 Plants SSH Backdoor — jd_pressman · 2026-10-30
- Stolen $1.45B in Bitcoin Is Nearly Unspendable: Blockchain Forensics Track Every Move — ayushtweetshere · 2026-10-30
- UK to pour another £120M into AI testing agency, outspending US counterpart 10x — DKokotajlo · 2026-10-30
- Serin: Researchers Demo a Backdoor Planted in an Open LLM — asankhs · 2026-10-30
- CrowdStrike: China-Linked Hackers Used DeepSeek LLM to Hit 200+ Orgs in 11 Days — BenBajarin · 2026-10-30
- GitHub Launches Preview of AI App Security Scanning for Prompt Injection and Agent Risks — sven_ai · 2026-10-30