CYBERSECURITYTRACKER
TRACKING6,506 stories in this site build1,309 vulnerability news stories in this site build
Permanent story citation

The Self-Expanding Stolen Inference Supply Chain: An AI Agent Harvesting and Re-Serving LLM Access

This page keeps the story as Cybersecurity Tracker first published it. If the tracker later corrects it, the correction appears below the original and never replaces it.

Back to newsStory 7095

As cited

Copy frozen at (site build).

ai security

The Self-Expanding Stolen Inference Supply Chain: An AI Agent Harvesting and Re-Serving LLM Access

A researcher documented an offensive operation where a semi-autonomous coding agent discovers poorly secured large language model (LLM) gateways, acquires access through web vulnerabilities and account farming, validates the stolen inference capacity, and consolidates it behind a unified gateway for reuse. The agent repeatedly located resale services using reconnaissance queries, created trial accounts with temporary email services, tested compromised credentials against premium models, and aggregated 379 upstream endpoints into a single New-application programming interface (API) instance serving five standardized model names. The attacker inadvertently exposed operational instructions and session context through honeypot captures, revealing how the agent edits databases to bypass rate limits and how the feedback loop creates a partially self-expanding inference supply chain.

Why it matters: LLM gateway operators and practitioners using untrusted or free LLM proxies face continuous agent-driven enumeration of authorization flaws, default credentials, and unauthenticated endpoints, while coding agents sending requests to malicious endpoints risk exposing operational state, source code, and command context alongside their prompts.

Source published
First seen by Cybersecurity Tracker

Source attribution

Glossary